Skip to main content

A quantum boost for machine learning

The blue room is dense with concentration. At a table in the centre sit two opponents staring at a board of black and white marbles that are moved in silent turns. Finally, the player on the right resigns. It is 9-dan Go master Lee Sedol. On the left sits software developer Aja Huang who gets his instructions from AlphaGo, a computer program developed by Google’s DeepMind project. It is March 2016 and AlphaGo has just beaten one of the world’s best players in four out of five matches of the popular board game Go.

The success of AlphaGo has been widely perceived as a milestone in artificial intelligence research. Go is a much more complex game than chess, at which a computer first won against a world champion in 1997. In Go, exploring all strategies by brute force – in which all possible moves are evaluated to decide the best move to make – is no option; there are more possible marble positions than there are atoms in the universe, and the 2200 computer processors delivering the power for the game are lightweight compared with today’s supercomputers. The secret of AlphaGo’s success lies much more in a strict training regime with a special sparring partner, namely the software itself. To become a worthy training partner, AlphaGo’s “deep neural networks” – computer algorithms inspired by our brain’s architecture – initially learnt how to master the game by consulting a database of around 30 million professional moves.

Machine learning can be understood as the data side of artificial intelligence, where one often deals with large amounts of information, or “big data”. Similarly to human learning, machine learning involves feeding very many instances of a problem into a computer that has been programmed to use patterns in the data to solve a previously unseen instance. For example, a computer could be fed a lot of images of a particular person, and then given a new image before being asked whether it is the same person as before. The crux is that we do not know how we link the visual stimulus to the concept of recognizing a person in the image. In other words, there is no simple correlation between the pixel at, say, position (1334, 192) being red and the picture containing our friend Sivu that we could programme the computer to exploit. Machine-learning research therefore has to come up with generic ways to find complicated patterns in data, and as Facebook’s automatic tagging function shows, this is done with ever-increasing success.

What does this have to do with physics, and more precisely, with quantum physics? The computer that executed the AlphaGo software is based on classical physics. Information is processed by microelectronic circuits that manipulate signals of zeroes and ones, and these circuits follow the laws of classical electrodynamics. But for two decades, physicists have been rethinking the concept of a computer right from scratch. What if we built a computer based on the laws of quantum theory? Would such a device fundamentally change the limits of what is computable? The answer, it turns out, is not so easy to find, although we seem to understand the question a lot better today. Despite the fact that we haven’t yet been able to build quantum computers large enough to solve realistic problems, several powerful languages have been developed to formulate and study “quantum algorithms”, the software for quantum computers, from a theoretical perspective. This research effort has now left the borders of purely academic interest and is pursued in the labs of large IT companies such as Google and IBM. As its realization seems more and more certain, the pressure to find “killer apps” for quantum computing grows. This is where machine learning comes in.

Since we know the language quantum computers will speak one day, we can already start thinking about what impact they will have on the frontiers of machine learning. This approach is called quantum-enhanced machine learning and is part of the larger research field of quantum machine learning (which also investigates the opposite approach of using traditional machine learning to analyse data from quantum experiments). To get an idea of quantum-enhanced machine learning, one first has to understand how machine learning works, and the “black art” involved in using it to greatest advantage.

Machine learning

A quick way to access the concept of machine learning is through data fitting, an exercise that most scientists have come across during their undergraduate studies and forms one of many methods used to recover patterns or trends in data. Imagine you run an experiment that generates data points (x, y) for setting a control parameter x and measuring the result y. As a physicist you would like to obtain a model that can explain these measurement results. In other words, you want to find a relationship y = f(x) that, up to some degree of noise, produced the data. This can be done by feeding the experimental data into a computer and using numerical software to find the best fit of a parameter-dependent function f(x) (figure 1). Mathematically speaking, this is an optimization problem.

A graph is marked with y as the vertical axis and x as the horizontal. Eight data points are marked on the graph as black crosses and labelled in the legend as "data". A blue dashed line, labelled in the legend as "model 1", passes through all the points and has several maxima and minima. A red dashed line, labelled in the legend as "model 2", doesn't pass directly through any points but goes approximately between them, having only one maximum and two minima. A black cross with a circle around it is labelled in the legend as "unseen data", and on the graph sits roughly in the middle horizontally, and below both plots vertically. A blue cross with a circle around it is labelled in the legend as "prediction of model 1", and sits way above the unseen data point, on the blue dashed line. A red cross with a circle around it is labelled in the legend as "prediction of model 2", and sits near the unseen data point, on the red dashed line

Solving the optimization problem is already the job half done for machine learning, where one can now use the best model function to predict the measurement outcomes for new control parameters without performing the actual experiment. Of course, in most machine-learning applications one is less interested in physical experiments than in tasks that traditionally require human experience. For example, x could represent a set of macroeconomic variables and y stand for the oil price development in the next week. If we derive a model y = f(x) from the data, we can use it to predict tomorrow’s oil price. Alternatively, the inputs could be the pixels of images and the output a yes-or-no answer to whether your friend Sivu is in the picture, in which case the machine-learning software is used for image recognition. One thing most applications have in common is that they allow us to answer questions about complex relationships where the answers are worth a lot of money.

So far this sounds pretty straightforward. All you have to do is solve an optimization problem to find the best predictive model. But machine learning usually deals with very difficult types of optimization problems that are avoided by even the more adventurous mathematicians. Think, for example, of an optimization landscape like the Himalayan mountain range, where you want to find the deepest valley on foot and without a map (figure 2). The real “black art” lies in the subtleties of formulating the optimization problem. In the data-fitting case of figure 1, for example, if we define the best model to be the one where the prediction f(x) is closest to the real value y for all data points, the more flexible model function (blue) would win, because the model function goes through all data points. But when we introduce a new data point, it is clear that the “rougher fit” (red) gives a much better prediction. For our hiker, the optimization landscape that corresponds to the more flexible model is not very helpful, because even if they find the deepest valley it does not necessarily lead to a good model. A useful optimization landscape leads to optimal models that generalize from the underlying pattern to unseen data, even if they do not predict the seen data perfectly well. Formulating effective optimization problems requires a lot of intuition and hands-on experience, which are key to harnessing the power of machine learning.

Upon a blue undulating 3D surface with gridlines on it resembling a mountain range, a white dashed line meanders this way and that from a white flag at the bottom of the hills to a red flag at the top

A quantum boost

The most common approach to enhancing machine learning with quantum computing is to outsource the hard optimization problems to a quantum computer, either the small-scale devices available in the labs today, or the full-blown versions that we hope to have access to in the future. An entire toolbox of algorithms for this purpose has been developed by the quantum-information-processing community. The continuing challenge is to combine, adapt and extend these tools with the aim of improving number crunching on conventional computers. Three different approaches to solving optimization problems using quantum computers are explained in more detail in the box opposite. Although we know by now that most hard computational problems tend to remain hard even if quantum effects are exploited, modest speed-ups can still prove crucial for today’s big-data applications.

There is one important caveat in outsourcing, however. For this approach to work, one needs to encode the data that shape the optimization landscape into the quantum system. One way to do this is to represent a black-and-white image with a lattice of spins pointing up (white pixel) or down (black pixel), as in figure 3. Using quantum superposition – where a physical system is in two or more quantum states at the same time – allows us to store many images in a single quantum system. Other encoding strategies are more involved, but all of them require in practical terms that we prepare the initial state of the quantum system (or, in some cases, the interactions) to represent the values of the dataset. For an experimental physicist, preparing a microscopic physical system so that it encodes billions of pixels from an image dataset, to a high precision, must sound like a nightmare. Data encoding is therefore a crucial bottleneck of quantum algorithms for machine learning and a challenge with no direct equivalent in classical computing.

A five-by-five grid of squares is shown sideways-on. Some are pale grey and some are dark grey, so that together as pixels they show the letter A. In every pale-grey square is a red arrow pointing upwards, and in every dark-grey square is a red arrow pointing downwards

Towards a quantum AlphaGo?

Without question, there is a long road to walk before future generations of AlphaGo and its companions can run on quantum hardware. First we need robust large-scale quantum computers to run the software developed. We need to design an interface between classical data and quantum systems in order to encode the problems in these devices. We also need better quantum tools for optimization, especially when the landscapes are complex.

More than anything, we need to learn the “black art” of machine learning from those who have been practising it for decades. Instead of merely outsourcing optimization tasks formulated for classical computers, we should begin to formulate problems for quantum computing right from the start. An early generation of quantum devices is waiting in the labs for practical implementations. The question is, what type of optimization problems do these devices allow us to solve, and can the answer to this question be used to define new machine-learning methods? Could there be specific physics-research problems that quantum-enhanced machine learning could tackle? Can we use genuine “quantum models” for these tasks? And can the way we think in quantum computing give rise to innovation for conventional strategies in machine learning?

In summary, the emerging discipline of quantum-enhanced machine learning has to be relocated from the playgrounds of quantum computing and must become a truly interdisciplinary project. This requires a fair bit of communication and translation effort. However, the languages might be less far apart than we expect: both quantum theory and machine learning deal with the statistics of observations. Maybe we do not need to take the detour via digital bits of zeros and ones in the first place. All in all, we do not know yet if some decades into the future, a quantum computer will calculate AlphaGo’s decisions. But asking the question gives us a lot to think about.

Approaches to quantum-enhanced machine learning

Quantum search
In the mid-1990s, computer scientist Lov Grover showed that a future quantum computer can search an unsorted database – such as telephone numbers in a phone directory – faster than classical computers can. This method can be adapted to find the k “closest” entries, or the k phone numbers that have the most digits in common with a given number. Finding closest data points to a new input is an important task in machine learning, for example in a method called “k-nearest neighbour” that chooses the new y-value according to the neighbours’ y-values. Maybe the most straightforward approach to machine learning with quantum computers is therefore to reformulate search problems in the language of quantum computing and apply Grover’s well-studied algorithm.

Linear algebra
A small quantum system can have a large number, N, of different configurations or measurement outcomes. Quantum theory describes the probability that one of the possible outcomes from all of these configurations is measured, and it is largely based on the mathematical language of linear algebra. In 2009 Aram Harrow, Avinatan Hassidim and Seth Lloyd from the Massachusetts Institute of Technology proposed a quantum algorithm that uses these properties in a clever way to solve systems of linear equations, which can under very specific circumstances be done incredibly fast. Likewise, many machine-learning optimization problems can be mathematically formulated as a linear system of equations where the number of unknowns depends on the size of the dataset. For big-data applications, numerical solutions can take a lot of computational resources and they are therefore excellent candidates for the application of the quantum linear systems algorithm.

Finding the ground state
A third type of optimization problem minimizes an overall energy function to find an optimal sequence of bits. A popular numerical method to carry out such “combinatorial optimization” is “simulated annealing”. Such an approach simulates the process in thermodynamics in which a system cools down until it reaches its ground state. In the quantum equivalent of the process, “quantum annealing”, an energy landscape is similarly minimized, however the algorithm can in addition use quantum-mechanical tunnelling to travel through the energy peaks – rather than having to “climb” over them – meaning it may find the lowest valley more quickly. Quantum annealing devices already exist, such as that announced in 2010 by the Canadian firm D-Wave as the world’s first commercial quantum computer. These devices have been shown to solve (admittedly, rather exotic) problems where they find a global minimum 100 million times faster than a conventional computer running a simulated annealing algorithm.

Coding and computing: the March 2017 issue of Physics World is out now

PWMar17cover-200By Louise Mayor

Physics these days wouldn’t succeed without software. Whether those lines of code are used to control new apparatus, make sense of fresh experimental data or simulate physical phenomena based on the latest theories, software is essential for understanding the world. The latest issue of Physics World, which is now live in the Physics World app for mobile and desktop, shines a light on how some physicists are exploiting software in new ways, while others are reinventing the hardware of a computer itself – binary isn’t the only way to go.

Sometimes there are so much data that software collaboration is the best way forward. In the issue, physicists Martin White and Pat Scott describe how the GAMBIT Collaboration is creating a new, open-source software tool that can test how theories of dark matter stack up against the wealth of data from various experiments such as direct searches for dark matter and the Large Hadron Collider. And with software development being so essential for physics research, data scientist Arfon Smith argues that we need to adopt better ways of recognizing those who contribute to this largely unrewarded activity. Columnist Robert Crease explores the other extreme: whether software can be patented.

Meanwhile, in an emerging field straddling both coding and computing, researcher Maria Schuld explains how quantum computers could enhance an already powerful software approach known as machine learning. (You can also read her article on physicsworld.com here.) Further into the realm of raw computing, physicist Jessamyn Fairfield describes the quest to develop a new kind of hardware that is physically, and functionally, similar to the computers inside our very own heads. As for how our brains process information, don’t miss a glimpse into the mind of physicist Jess Wade who has created a doodle based on the work Fairfield describes.

(more…)

Supernova 1987A enters a new phase

Thirty years after it exploded, supernova SN 1987A is starting a new phase in its development as the shock wave from the stellar explosion is finally passing beyond a ring of gas encircling the dead star.

On 23 February 1987, a blue supergiant star named Sanduleak –69° 202 exploded in the Large Magellanic Cloud, which is a dwarf-galaxy neighbour to the Milky Way 169,000 light-years away. Named SN 1987A, it was the first supernova since 1604 to be visible to the naked eye.

Robert Kirshner, an astrophysicist at the Harvard–Smithsonian Center for Astrophysics in the US, told physicsworld.com that “SN 1987A is unique” in terms of the unprecedented scrutiny it has received as the nearest supernova in the modern age. Observations made over the past 30 years are teaching us what happened to Sanduleak –69° 202. And by comparing the light given off by SN 1987A to that from more distant supernovae we can learn about these objects and their progenitor stars.

Glowing rings

The flash of light from SN 1987A was preceded by a burst of neutrinos that arrived at Earth 3 h before the visible light. Light from the explosion illuminated three rings that surround the supernova. The two outer rings are faint and distant, but the inner ring is dense and clumpy, with a diameter of about a light-year. They are all made from material emitted from the star as it underwent pulsations in its outer layers tens of thousands of years before it exploded, therefore providing us with a window through time to show how the star was behaving in the run-up to its destruction.

After the initial flash of light from the supernova the rings faded, only for the inner ring to brighten once again when the supernova shock wave caught up with it in 2001. The shock caused the inner ring to heat up and emit X-rays, detected by NASA’s Chandra X-ray Observatory, with energies mostly in the realm of 0.5–2 keV, but at its peak as high as 8 keV. The inner ring continued to brighten until 2013, at which point it began to fade in uneven fashion as a result of being shredded by the expanding blast wave moving at 1800 km/s.

Intriguingly, the inner ring is fading unevenly, but is the ring itself lopsided “or does the uneven interaction of the shock with the ring indicate that the supernova explosion itself was asymmetric”? This is a question asked by Kari Frank of Penn State University, who led the most recent Chandra observations of SN 1987A. If the ring is lopsided then it could be the result of Sanduleak –69° 202 being part of a binary system, with the gravity of an unseen companion star influencing how the ring material was thrown into space.

Missing pulsar

There’s also the question of what the supernova left behind. Sanduleak –69° 202 had a mass estimated to be 20 times greater than the Sun and should have created a spinning neutron star – a pulsar – when its core collapsed and emitted the neutrino burst. Yet so far there has been no evidence that a pulsar is present. It could be that its beams are pointing away from us, but a pulsar should also produce thermal X-rays from its hot surface as well as a wind of radiation, neither of which has been observed.

“The most likely reason we have not seen any of this yet is because there is a lot of cold gas and dust still hanging out near the centre of the ring,” says Frank. Like a thick fog, this cold gas and dust could be blocking the pulsar’s emissions but, as that fog expands along with the rest of the supernova remnant, it will thin and eventually dissipate, revealing the pulsar within.

The reveal of the pulsar is one event that could transpire in the next 30 years, a time during which SN 1987A will “transition from a supernova to a supernova remnant, shaped and powered by the collision of the shredded star with the surrounding gas,” says Kirshner.

Cooling dust

Supernovae remnants are characterized by the cooling of dust spewed out into space by the stellar explosion. This dust contains elements such as carbon, oxygen, nitrogen, silicon and iron, all forged within the dead star. The cooling dust is visible to the Atacama Large Millimeter/submillimeter Array (ALMA) in Chile, which will monitor how the supernova is dispersing the dust into space to be recycled in the next generation of stars, planets and possibly even life.

Consequently, the next 30 years, like the previous three decades, will be a process of learning as astronomers begin to join the dots between SN 1987A’s remnant and other young supernova remnants in the Milky Way. There will undoubtedly also be plenty more surprises as the shock wave moves into new territory beyond the inner ring.

What will we find there? “We’re about to find out,” says Frank.

Flash Physics: Frogs see colour at night, new lens sharpens X-rays, PICO-60 puts new limits on dark-matter

Amphibians have colourful night-time vision

Frogs and toads can see colour when it is too dark for humans to see. Many animals have evolved to have impressive visual skills, yet understanding how and what they see has always been a challenge for scientists. A group from Lund University in Sweden and the University of Helsinki in Finland has therefore used behavioural studies to investigate the vision of frogs. Human eyes contain two types of visual cells in the retina – cones and rods. The cones allow us to see colour, but they stop working at night because they require a lot of light. At this point the rods take over and enable us to see in black and white under low-light conditions. Although this is the case for most vertebrates, frogs and toads are unique in that they have two spectrally different kinds of rod photoreceptors. Scientists have long hypothesized that these allow them to see in light levels that are too low for humans. While Carola Yovanovich and colleagues have shown this is true, they have also found that the amphibians can actually see colour in extreme darkness. The team performed three behavioural experiments to study how the frogs and toads used vision during different natural scenarios in near complete darkness. When searching for a mate, the animals stopped using colour information fairly early on in the process. Meanwhile, when hunting for food and escaping dens and passageways, they used all sensory information available to them including the ability to see colour in extremely dark surroundings. The study can be found in the Philosophical Transactions of the Royal Society B theme issue on “Vision in dim light”.

New lens sharpens X-ray beams

Profile of a focused X-ray beam

X-rays can now be focused with much greater precision using a new corrective lens developed by an international team of scientists. The component is designed for use on X-ray synchrotrons, which supply coherent beams of radiation for applications including condensed-matter physics, biology and chemistry. X-rays obey the same optical laws as light but because they have very short wavelengths and much higher energies it is very difficult to make lens systems that can steer and focus them. In particular, lens distortions of just a few hundred nanometres can have a detrimental effect on X-ray optics. While high-quality lenses made from beryllium are available, better optics could lead to improved measurements and even new types of experiments. Now, researchers at several universities in Germany and Sweden, and synchrotron labs in the UK, US and Germany, have carefully measured the distortions in a stack of beryllium lenses. They then used this information to create a corrective lens that was milled using a precision laser. Without the new lens, the beryllium stack could focus X-rays to a spot about 1600 nm in diameter. This was reduced to 250 nm by using the corrective lens. The research is described in Nature Communications.

PICO-60 puts new limits on dark-matter interactions

Photograph of the PICO-60 detector

Physicists working on the PICO-60 detector in Canada have put a new upper limit on the strength of the spin-dependent interaction between the proton and hypothetical dark-matter particles called WIMPs. Located 2 km underground to shield it from cosmic rays, PICO-60 is one of several experiments running at SNOLAB in Sudbury, Ontario. It contains 52 kg of the liquid octafluoropropane, which is maintained in a superheated state. If a WIMP collides with a fluorine nucleus, the energy imparted to the fluid will cause it to boil locally, creating a bubble that can be detected using an array of digital cameras and acoustic detectors. While no dark-matter interactions were spotted, the experimental run allows physicists to put new upper limits on the strength of the spin-dependent interaction between WIMPs and protons at WIMP masses between about 10–100 GeV/c2. This will help to guide future dark-matter searches, and the PICO team is now developing a larger version of the detector that will use 500 kg of superheated fluid. The latest results are described on arXiv.

 

  • You can find all our daily Flash Physics posts in the website’s news section, as well as on Twitter and Facebook using #FlashPhysics. Tune in to physicsworld.com later today to read today’s extensive news story on how a supernova is entering a new phase of life.

Flash Physics: Flower acts as supercapacitor, ORNL head steps down, triboelectricity boosts mass spectrometer

Flower doubles as supercapacitor

Two years ago, researchers at Linköping University in Sweden showed that a rose can form the basis of a transistor. Now team member Roger Gabrielsson and colleagues have used a similar flower to create a supercapacitor that can store a large amount of electrical energy. The team put a cut rose into a polymer solution, which was absorbed by the flower. The material polymerizes spontaneously within the stem, leaves and petals of the flower to create long threads that conduct electricity. This allows a large amount of electrical change to be pumped into the flower. “We have been able to charge the rose repeatedly, for hundreds of times without any loss on the performance of the device,” explains team member Eleni Stavrinidou. “The levels of energy storage we have achieved are of the same order of magnitude as those in supercapacitors.” She adds that without any further optimization, the rose supercapacitor is capable of powering an ion pump and various types of sensors. The rose supercapacitor is described in the Proceedings of the National Academies of Science.

Head of Oak Ridge National Lab steps down

Photograph of Thom Mason

Thom Mason has announced he will step down as director of Oak Ridge National Laboratory (ORNL) on 1 July 2017 – exactly 10 years after first taking the job. ORNL is the US’s largest science and energy laboratory and focuses on materials, neutron science, energy, high-performance computing, systems biology and national security. It operates two neutron facilities – the High Flux Isotope Reactor and the $1.4bn Spallation Neutron Source – as well as two supercomputers. Mason, an experimental condensed-matter physicist originally from Canada, will become senior vice president for laboratory operations at Battelle – a private non-profit science and technology firm that is based in Columbus, Ohio. Battelle, together with the University of Tennessee, manages ORNL for the US Department of Energy. ORNL is now looking for Mason’s replacement.

Triboelectric generator gives mass spectrometer a boost

Photograph of Anyin Li using a sliding triboelectric nanogenerator

A device that generates electricity from friction has been used to increase the sensitivity of a mass spectrometer to “unprecedented levels”, according to Facundo Fernández, Zhong Lin Wang and colleagues at the Georgia Institute of Technology in the US. Triboelectric nanogenerators (TENGs) convert mechanical energy to electrical energy and are of great interest to Wang and others who are building systems to harvest energy from the environment. The team’s TENG provides an ionizing voltage of 6000–8000 V to a mass spectrometer, compared with the 1500 V that is supplied by a conventional source. The TENG voltage alternates in polarity and the corresponding current is supplied in a very precise manner. The team believes that this voltage is extremely efficient at ionizing the sample – the first stage of operation of a mass spectrometer – and this means that a stronger measurement signal is obtained. The alternating and controlled nature of the signal also allows higher voltages to be used without damaging the sample or the mass spectrometer. Most mass spectrometers operate in a pulsed mode, so a pulsed TENG ionization voltage could be synchronized to ensure that the sample is only ionized when necessary. This could allow very small samples to be analysed at higher sensitivities because precious ions would not be wasted. Fernández says that the TENG allowed “us to reach sensitivity levels that are unheard-of – at the zeptomole scale”. TENGs could also make mass spectrometers more compact and portable by replacing conventional ionization voltage supplies, which tend to be large and bulky. Looking further into the future, it may even be possible to create mass spectrometers that are powered by TENGs alone. The research is described in Nature Nanotechnology.

 

  • You can find all our daily Flash Physics posts in the website’s news section, as well as on Twitter and Facebook using #FlashPhysics.

Speaking a different language: how to communicate science

From books and lectures to blogs and social media, today’s physicists are expected to do far more public communication than their predecessors. However, communicating science well is far from easy. Obstacles include the comfortable familiarity of jargon and a natural distaste for the imprecision required by simplification. But there is a more fundamental barrier, which arises from scientists and non-scientists speaking different languages and having different mental models of how the world works.

In my career as a science communicator, I have observed countless online exchanges between scientists and the public, I’ve read and reviewed more than 1000 popular-science books, written 26 of my own, and spoken to numerous readers and attendees at public lectures. My conclusion is that physicists find communication hard because their idea of the questions physics is trying to answer is so different from those of their readers and the wider public. Although physicists aren’t unique in these problems, the concepts they need to communicate are often less immediately accessible than those from biology, chemistry and other fields.

“Why” versus “how”

The scientist–public language barrier was perfectly highlighted in a blog post written last September by Sabine Hossenfelder, a theoretical physicist and research fellow at the Frankfurt Institute of Advanced Studies (ow.ly/NsJO3077wZc). Hossenfelder puts a lot of effort into communication, not only through her Backreaction blog but also interacting via social media and writing for magazines, including Physics World (see “Can we unify quantum mechanics and gravity?” October 2013 pp42–43). She is anything but an ivory-tower physicist.

In her post, Hossenfelder analyses a question-and-answer discussion on Discover magazine’s website last September between staff editor Bill Andrews and a reader called Jeff Lepler from Michigan in the US. “Are we,” wonders Lepler, “any closer to understanding the root cause of gravity between objects with mass?” To which Andrews replies: “Sorry, Jeff, but scientists still don’t really know why gravity works. In a way, they’ve barely just figured out how it works.”

To the public, science is about the search for simple truths that explain how nature functions and, as a result, possibly also getting some idea of the why

In response, Hossenfelder asks “What’s that even mean – scientists don’t know ‘why’ gravity works?” She asks this rhetorically, going on explain that why gravity works is not a scientifically meaningful question; it’s rather how gravity works that physicists aim to answer.

It has only been since the development of modern science, however, that scientists have concentrated on answering how something works. In the time of natural philosophy – the precursor to modern science – scholars did indeed attempt to answer the question why. For this reason, some modern scientists write negatively in their popular books about the physics of, say, Aristotle and his contemporaries. But the natural instinct of Ancient Greek philosophers – shared by many people today – was to be driven by the search for causal reasons. Aristotle’s explanation of gravity depended, in other words, not on how things were attracted to the centre of the universe, but rather why this happened.

Aristotle believed that the four elements (earth, air, fire and water) each had their own natural tendencies. The reason why, for instance, an Earth-based object like a rock fell, he argued, was because its natural place was the centre of the universe (which meant the Earth). Aristotle tried to explain why the rock fell, not how it fell. Modern scientists can easily dismiss the philosopher’s explanation as irrelevant. But if they don’t understand the thought process behind the question, they miss a natural line of thinking for a non-scientist, which needs to be addressed rather than dismissed.

More fundamental for Hossenfelder than the language niggle, though, was the suggestion that scientists have “barely figured out how [gravity] works”. In her blog post, Hossenfelder responds that we do know how gravity works. “The purpose of science is to explain observations,” she writes. “We have a theory by the name General Relativity that explains literally all data of gravitational effects. Indeed, that General Relativity is so dramatically successful is a great frustration for all those people who would like to revolutionize science à la Einstein. So in which sense, please, do scientists barely know how it works?”

But answering the question “How?” itself merely generates a second-level question. If you said gravity works by warping space–time, the immediate response of a non-scientist would be: “Yes, but how does matter warp space–time?” And now physicists are stumped. They can’t answer that question, because repeatedly asking “How?” leads to a dead end.

Physicists may describe constructing a mathematical model as understanding, but it isn’t what most people mean by the word

To the public, science is about the search for simple truths that explain how nature functions and, as a result, possibly also getting some idea of the why. But when a physicist talks of understanding something, they have in mind being able to construct a model – usually mathematical – that matches, as closely as possible, the data that are observed. Physicists may describe this as “understanding”, but it isn’t what most people mean by the word. I feel it’s this lack of a shared vision between the public and the physics community that creates an inevitable barrier.

Offering lettuces to an ass

Galileo and Newton

It can sometimes feel to the public that scientists are guilty of perpetuating the concept that wisdom should be kept from the masses. Writing in the 13th century, friar and proto-scientist Roger Bacon noted that Aristotle said “It is stupid to offer lettuces to an ass since he is content with thistles.” Bacon agrees, telling us that “The cause of obscurity in the writings of all wise men has been that the crowd derides and neglects the secrets of wisdom and knows nothing of the use of these exceedingly important matters.”

This philosophy comes through in Isaac Newton’s Principia, particularly when set alongside Galileo’s Two New Sciences. Each of these books is a masterpiece of physics. Yet the approach taken could not be more different. Galileo writes in his native language, Italian. His prose is accessible and his discussions of physical principles take the form of a three-way discussion between two experts (Salviati and Sagredo) and a third person, Simplicio. Presenting the voice of the ordinary man, Simplicio is the most interesting participant in this context. He asks the questions ordinary readers want answered, a simple device that enables Galileo to keep his text grounded.

By contrast, Newton writes in Latin, making his book less accessible to the non-expert. Indeed, Newton goes out of his way to make the book difficult to read. He notes that he originally intended the third part of the book, The System of the World, to be for a popular audience “so that it might be more widely read”. However, he then deliberately recast it to make it more obscure.

“Those who have not sufficiently grasped the principles set down here,” he remarks, “will certainly not perceive the force of the conclusions, nor will they lay aside the preconceptions to which they have become accustomed over many years; and therefore, to avoid lengthy disputations, I have translated the substance of the earlier version into propositions in a mathematical style, so that they may be read only by those who have first mastered the principles.”

Perhaps physicists should be a little more like Galileo and a little less like Newton when communicating with the public.

Delivering the goods

Getting a better feel for what the public’s questions mean isn’t the only prerequisite for good communication. Another danger is that if we take too much for granted, our explanations don’t deliver. In a recent book, A Farewell to Ice (2016 Allen Lane), Peter Wadhams – a sea-ice scientist and former professor of ocean physics at the University of Cambridge – tries to explain the concept of Fourier analysis to the reader. “The Fourier series,” he writes, “by which any function can be split into a set of harmonics…” – without thinking that anyone who knows what a function being split into harmonics involves probably knows about Fourier analysis. It’s a non-definition – a trap that many communicators fall into.

However, there is a further barrier when it comes to understanding and anticipating the questions the public might ask. When scientists are writing for the public they need to bear in mind the kind of questions that their words will generate in the mind of the reader and be prepared to answer those questions. Paradoxically, whereas using too many technical terms over-complicates an explanation, a lack of anticipation leads to over-simplification.

As an example, in his otherwise excellent book Neutrino Hunters (2013 Scientific American), Ray Jayawardhana – an astrophysicist at York University in Canada – describes the challenges facing people who build detectors for these elusive particles. Jayawardhana introduces us to a development of the Japanese Super-Kamiokande detector: “Beacom and his colleagues have suggested that dissolving a bit of gadolinium, a silvery-white metal, in the giant water tank at Super-Kamiokande would do the trick [of distinguishing supernova relic neutrinos], since the fix would enhance the detector’s sensitivity to relic neutrinos.” Then he moves on. Gadolinium is not mentioned again.

In that extract, we see Jayawardhana answering a potential reader question – “What is gadolinium?” – though, to be frank, giving the element’s colour feels a bit like asking what kind of car someone has and being told that it’s blue. However, the author leaves the reader mentally stranded. Jayawardhana knows the answers, but doesn’t share them. As a result, the reader is mired in questions that are never answered. How does gadolinium improve detection? How does this enable the detector to distinguish relic neutrinos from common-or-garden local ones? We don’t find out.

What is lacking here is an ability to pre-empt questions that the reader may have in response to a piece of writing, and then to formulate a response with which the lay person will be satisfied. Let me use one more example to show this barrier in action.

Mind the warp

A few years ago, I was researching my popular-science book Gravity (2012 St Martin’s Press). In the section on the general theory of relativity, I had to get across how it is that the warping of space–time produced by matter can lead to the effects that we experience as the force of gravity. Inevitably I wheeled out the “bowling ball on a sheet of rubber” analogy, which uses a flat sheet of rubber, stretched taut, as a 2D model of space.

Representing the effects of gravity

We imagine a beam of light, or the straight-line motion of a planet flying free through space, as a coloured line on the surface of the sheet. We then place a bowling ball on the sheet and it distorts the rubber. As a result, the line is no longer straight. It curves around the ball. This, we say, is rather like the way that matter distorts space (though in 3D, rather than 2D), causing light and the straight-line motion of celestial bodies to bend around massive objects like stars and planets. At this point, traditionally, the popular account of gravity moves on. But there is a question that is not asked. To take Newton’s example: how does this make an apple fall?

If this is ever mentioned when using the rubber-sheet analogy, we usually get some hand-waving suggestion that the apple slides down the indentation in the rubber sheet towards the massive object. But what makes the apple move? Gravity – so we haven’t got anywhere. Because I couldn’t come up with a sensible extension to the rubber-sheet analogy, I e-mailed a wide range of physicists asking how they would explain, for the general public, how the apple goes from being stationary to falling.

Most academics didn’t respond. This might seem unsurprising, but, on the whole, professional scientists are happy to answer meaningful queries. The apparent implication here is that either the rubber-sheet analogy has been stretched too far or that the answer is just too complicated for little minds to worry themselves with. Where I did get replies, they tended to suggest my question was irrelevant or, well, the mathematics works, so let’s not worry about the analogy. A typical argument from those in the “irrelevant” camp” was: “A force is a force, of course, and asking why/how it causes an acceleration already seems a bit circular.” Asking how you can have a force producing action at a distance is not at all circular, but something that has worried people for a long time. There still needs to be a cause.

Finally, I got an explanation I could use from Fried­rich Wilhelm Hehl of the University of Cologne in Germany. His initial explanation, admittedly, had limited value for the public: “As you correctly say, in the ‘bowling ball on a rubber sheet’ picture, the sheet represents space–time. In space–time the 4-velocity of a particle is just the tangent unit-4-vector of the path traced by the particle. Since a curve in the sheet at a certain point has always a tangent 4-vector, the particle has to follow the path with the corresponding 3-velocity.”

This was useful to know, I’m sure, but didn’t exactly illuminate things for the non-physicist. But after a little probing, Hehl translated this explanation into a more amenable form. By getting the reader to think not of distorting space but space–time, and thinking through the implications of warping the time dimension, which with nowhere else to go has to produce a change in the space dimensions, it was possible to extend the analogy to see how warped space–time could produce motion.

Lessons to learn

Altogether, I’d suggest that there are three significant lessons from these examples. First, if you get a question that doesn’t make sense to you, don’t simply assert that it is a silly question. More likely, it’s the scientist who is at fault for not understanding – not the questioner for being dumb. Ask for more detail. And if you still don’t get it, ask a friend who isn’t a physicist what it means. (If you don’t have any friends who aren’t physicists, get out more.) Ensure your answer takes into account the questioner’s viewpoint.

Try to break down what you are saying and identify where the assumptions are

Second, when you are explaining something, always try to be aware of your assumptions about what the questioner already knows. This can be difficult, because so much of what you assume has been part of your working life for so long. Try to break down what you are saying and identify where the assumptions are – then make sure that they are ones that hold for your audience too. If not, deal with them.

And finally, don’t over-simplify. We are used in academic writing to ensuring that statements are backed up by sources or experimental evidence. In public communication, it’s important to back up statements with explanations that the reader needs (and not just the explanations that you think you need). Provided you do explain appropriately, you can go into more detail than you might think. Many books on gravity for a lay audience don’t include Einstein’s field equations. Mine did – not because I felt the reader is ever going to do the maths, but to satisfy their curiosity of what these things look like and to explain the main elements of the equations, without straying into mathematical complexity.

If physicists are to communicate well with the public, we need to do more than just weed out jargon and simplify complex mathematics. We also need to understand the nature of the public’s questions, anticipating queries that are based on a different kind of thinking. It’s a stretch, but it is possible – and, I would argue, is essential if we want the public, which controls the purse strings, to support progress and advancement in physics.

Climate-friendly aircraft routing could cut environmental damage

Rerouting transatlantic flights to follow the most climate-friendly path could damage the climate 10% less for an increase in costs of just 1%. That’s according to a team from Germany, the Netherlands, Belgium, Norway and the UK.

“An attractive aspect of our approach is that it potentially enables some mitigation of aviation’s climate impact…using the current aircraft fleet,” Volker Grewe of the Deutsches Zentrum für Luft- und Raumfahrt, Germany, and Delft University of Technology in the Netherlands, explains. “Some mitigation options involve changes in aircraft or engine design, which would take decades to implement given the slow – and expensive – turnover of the global fleet.”

Volker and colleagues modelled routings for 800 daily flights across the Atlantic under five typical winter weather patterns and three typical summer patterns. The team combined the EMAC chemistry-climate model with an air-traffic simulator, choosing 85 variations for each flight path – 17 horizontal and five vertical. Then they picked the most “eco-efficient”, which is the path with the best ratio of climate-impact reduction to cost increase.

Multiple impacts

Aircraft have an impact on the climate by emitting carbon dioxide, water vapour, nitrogen oxide and particulates. These alter the concentration of the greenhouse gases ozone and methane, and also form contrails. Where and how high the plane is, as well as the time of day and season, all alter the size of its climate effect.

“It is now well established that – unlike the climate effects of CO2 – the non-CO2 climate effects such as contrail formation depend sensitively on when and where the aircraft emissions occur,” says Grewe, “and these sensitive regions vary in location and importance from day to day, as they are strongly influenced by the prevailing weather patterns.”

Contrails, for example, form if the hot, moist exhaust from the jet engine becomes saturated with respect to water when it mixes with the air in the atmosphere. And the trails only persist if the ambient air is saturated with respect to ice. Contrails affect both incoming radiation from the Sun and the exit to space of infrared radiation emitted by Earth and its atmosphere. On average, the trails cause warming, but close to sunrise and sunset they can result in cooling.

Ozone and methane

In general, aircraft emissions tend to boost the amount of ozone and decrease methane concentrations, with the warming from the extra ozone outweighing the cooling from the reduction in methane. But this varies a lot locally, and in some regions the emitted nitrogen oxides cause cooling.

“Put simply, if we can avoid those regions in the atmosphere where the non-CO2 emissions have the largest climate effect, we can reduce the climate impact significantly,” says Grewe. “Our modelling study showed that a large reduction of aviation’s climate impact is feasible at relatively low costs.”

Rerouting flights could cut their climate harm but may increase fuel and staff costs. Cost-efficient reductions in climate impact mostly resulted from avoiding the formation of warming contrails or from producing cooling contrails, Grewe and his colleagues found.

Close collaboration

Investigating such mitigation options requires close collaboration between atmospheric scientists and disciplines like air-traffic management, Grewe says. He believes all sectors must play a role in meeting the internationally agreed 2 °C target for the total climate effect of human activity.

“This is particularly challenging for the aviation sector, given the predictions of its continued growth over coming decades,” he says. “Hence, we need a combination of various mitigation options – technological, such as cleaner and more efficient engines, and operational, i.e. more eco-efficient routes. To make this happen, a political framework is required, which aims at limiting aviation impacts. It might be on an international or regional basis.”

Airlines would in all likelihood need regulations or a market incentive such as a price on climate impact to carry out such climate-friendly routing. While the International Civil Aviation Organisation (ICAO) has decided to implement the Carbon Offsetting and Reduction Scheme for International Aviation (CORSIA), non-carbon-dioxide effects are still not considered in political decisions to limit the climate effect of aviation, Grewe says.

Reliable forecasts

“Implementing our proposed approach…is likely to be at least 5–10 years in the future – it should be considered ‘exploratory’ at present,” he adds. “We have to convince all stakeholders that the approach is worthwhile and feasible in practice, and that the costs associated with it are proportionate. And because the location of the climate-sensitive regions varies markedly from day to day, we also need to clearly establish that we can reliably forecast these areas sufficiently far in advance, so that re-routing aircraft to avoid them can be done with confidence.”

Now the scientists, who included a road map in their paper in Environmental Research Letters (ERL), are looking for more funding. They are also participating in ATM4E, a European project investigating whether it’s possible to avoid climate-sensitive regions in areas with high traffic density, as well as how the approach can be made operational, included in a weather forecast system, and verified.

New graphene-like material could have a band gap

A new 2D material just one atom thick has been made by an international team of researchers led by Axel Enders. Dubbed hexangonal boron–carbon–nitrogen (h-BCN), the material could offer many of the benefits of graphene, which is a hexagonal lattice made of just carbon. But unlike graphene, h-BCN has a direct electronic band gap, which could make it useful for creating electronic devices.

First isolated in 2004 by Andre Geim and Konstantin Novoselov, who shared the 2010 Nobel Prize for Physics for their discovery, graphene is blessed with a wealth of potentially useful mechanical and electronic properties. Despite being so thin and flexible, graphene is much stronger than steel and is an excellent conductor of heat. Graphene also conducts electrons at extremely high speeds. As a result, it could be the ideal material for making electronic devices that are ultrafast, high-density and even bendable.

Mind the gap

Many of graphene’s electronic properties arise from the fact that it is a semiconductor with a zero-energy gap between its valence and conduction bands. This is not ideal for making transistors and other electronic devices because such circuits need semiconductors, such as silicon, that have a band gap. In an attempt to make a modified version of graphene that does have a band gap, device developers have therefore explored various schemes – including applying an electric field, adding chemical impurities or modifying the structure of graphene. None, however, has proved ideal.

Now, Enders and colleagues at the University of Bayreuth, University of Nebraska-Lincoln, University of Krakow, State University of New York at Buffalo, Boston College and Tufts University have developed a graphene-like material that could fit the bill. The team made h-BCN by heating an organic molecule containing boron, nitrogen and carbon on an iridium substrate. The result is an atomically thin layer of h-BCN that is corrugated because of the lattice mismatch between the layer and substrate.

Multiple techniques

The team studied the structure and electronic properties of the film using molecular-resolved scanning tunnelling microscopy imaging, X-ray photoelectron spectroscopy and low-energy electron diffraction. The researchers also used density functional theory and first-principles calculations to further understand h-BCN.

An important result of the measurements and calculations is the prediction that h-BCN should have a direct electronic band gap of a size that falls between that of gapless graphene and hexagonal boron nitride, which is an insulator. According to the researchers, this band gap could make the material better suited than graphene for electronics applications.

“Our findings could be the starting point for a new generation of electronic transistors, circuits and sensors that are much smaller and more bendable than the electronic elements used to date,” says Enders. “They are likely to enable a considerable decrease in power consumption.”

The study is described in ACS Nano.

Flash Physics: Atoms mimic each other, QCD cracks five loops, metamaterial bricks shape sound

How atoms can impersonate each other

An atom could be made to emit an optical signal that is usually associated with another type of atom, according to calculations done by Andre Campos, Denys Bondar, Herschel Rabitz and Renan Cabrera at Princeton University in the US. When an atom is illuminated with light it can absorb energy and give off light at a set of frequencies distinct to that type of atom – which forms the basis of optical spectroscopy. However, if the atom is illuminated by an intense and complex optical signal it should be possible – in principle – to control the quantum states of the atom and cause the emission of light at frequencies not normally seen from that atom. Unlike conventional spectroscopy, a measurement of such a spectrum would not reveal the type of the atom – unless the experimenter knew the precise nature of the complex optical signal. Previous attempts to calculate the exact nature of such an optical signal has proven very difficult. But now, the team has come up with a successful scheme that involves both bound and ionized quantum states of an atom. Writing in Physical Review Letters, the team points out that some of the experimental techniques needed to carry out its scheme have already been demonstrated in the lab.

Five-loop QCD calculated at long last

A quantum chromodynamics (QCD) calculation involving five loops has been made for the first time by physicists in Russia and Germany. QCD describes the strong nuclear force between the quarks that make up protons, neutrons and other heavy particles. It is notoriously difficult to calculate the properties of systems governed by QCD because of the enormous strength of the strong nuclear force and the fact that calculations must consider large numbers of virtual quark–antiquark pairs that pop into and out of existence. As a result, physicists have struggled to calculate the properties of even simple objects such as the proton. Since the early 1970s, physicists have shown that QCD calculations can be made as a series of corrections to a leading-order calculation. These corrections are called loops, and physicists had been able to calculate one-, two-, three- and four-loop corrections. However, progress had been stuck at four loops since 1997. Now, Pavel Baikov at the Skobeltsyn Institute of Nuclear Physics in Moscow and Konstantin Chetyrkin and Johann Kühn of the Karlesruhe Institute of Technology have extended calculations to five loops. Writing in Physical Review Letters, the trio use five loops to calculate several properties of the Higgs boson.

Metamaterial bricks shape sound

Photograph of metamaterial bricks

A new “supermaterial” has been made that can bend and shape sound waves using specially designed bricks. Scientists at the University of Sussex and the University of Bristol in the UK have developed an acoustic device that can transform incoming sound waves into any required sound field. Sound manipulation is useful for many applications including ultrasound imaging, loudspeaker design and acoustic levitation. Current approaches use fixed lenses and expensive phased arrays. In contrast, this latest device comprises small, 3D-printed metamaterial bricks. These slow down incoming sound by directing it through meandering channels. The tailored geometries of the channels delay the wave phase to create the desired sound field. Gianluca Memoli from the Sussex team describes the device as a “do-it-yourself acoustics kit”, as the bricks are easily made and can be arranged in arrays specific to the application requirements. The new material could be used on a large scale to direct and focus sound to form an audio hotspot. It could also be suitable for small-scale applications such as focusing high-intensity ultrasound waves to destroy tumours within the body. The material is presented in Nature Communications.

 

  • You can find all our daily Flash Physics posts in the website’s news section, as well as on Twitter and Facebook using #FlashPhysics. Tune in to physicsworld.com later today to read today’s extensive news story on a new graphene-like material.

A laser-bubble mermaid, ode to seven exoplanets, metallic hydrogen is lost

Tiny bubbles: laser-made mermaid (Courtesy: Kota Kumagai, Utsunomiya University)

By Hamish Johnston

A popular way of melding science and art is to create an image of a mythical being in your lab. Yoshio Hayasaki and colleagues at Utsunomiya University in Japan have made a pretty good likeness of a mermaid using a laser that forms tiny bubbles inside a liquid. “In our display, the microbubble voxels are three-dimensionally generated in a liquid using focused femtosecond laser pulses,” explains team member Kota Kumagai.

(more…)

Copyright © 2026 by IOP Publishing Ltd and individual contributors