Tyler O'Neal, Staff Editor ECONOMICS

Japanese-built MD simulations, ML, topology reveals a hidden relationship in amorphous silicon

Theoretical scientists have used topological mathematics and machine learning to identify a hidden relationship between nano-scale structures and thermal conductivity in amorphous silicon, a glassy form of the material with no repeating crystalline order.

A study describing their technique appeared in the Journal of Chemical Physics yesterday.

Amorphous solids, such as glass, obsidian, wax, and plastics, have no long-range repeating, or crystalline structure, to the atoms or molecules that they are made out of. This contrasts with crystalline solids, such as salt, most metals, and rocks. As they lack long-range order in their structure, the thermal conductivity of amorphous solids can be far lower than a crystalline solid composed of the same material.

However, there can still be some medium-range order on the scale of nanometers. This medium-range order should affect the propagation and diffusion of atomic vibrations, which carry heat. The heat transport in disordered materials is of special interest to physicists due to its importance in industrial applications. The amorphous form of silicon is used in an enormous range of applications in the modern world, from solar cells to image sensors. For this reason, researchers have intensively investigated the structural signature of the medium-range order in amorphous silicon and how it relates to thermal conductivity.

“For better control over applications that make use of amorphous silicon, controlling its thermal properties is high on engineers' wish list,” said Emi Minamitani, the corresponding author of the study and a theoretical molecular scientist with the Institute for Molecular Science in Okazaki, Japan. “Extracting the nano-scale structural characteristics in amorphous including medium-range order is an important key.”

Unfortunately, researchers have struggled to carry out this task because it is difficult to determine the essential nano-scale features of disordered systems using traditional techniques.

In experiments, the presence of medium-range order has been physically detected using fluctuation electron microscopy, which involves statistical analysis of scattering from nano-scale volumes of a disordered material. At the theoretical level, it has been discussed by considering the distribution of dihedral angles (the angle between two intersecting planes between sets of atoms) or using ‘ring statistics.’ The latter tries to understand the structural characteristics from the connectivity of atoms.

This in turn draws on the field of mathematics known as topology, which investigates properties of an object that do not change—or are ‘invariant’— even when the object is constantly stretched and deformed without being broken (such as shapes written on a rubber sheet). Focusing on this topological invariance is useful for delivering a qualitative description, such as the tendency of the physical properties with respect to the randomness. However, it is demanding to determine the atomic structure corresponding to a medium-range order and predict its physical properties only from simple topological invariants.

So the researchers pivoted to an emerging technique called persistent homology, a type of topological data analysis. Persistent homology has been used elsewhere to analyze complex structures ranging from proteins to amorphous solids. The benefit of this method is in detecting topological features in complicated structures at different spatial scales. This is vital because the medium-range order comprises quasi-repetitive structures at various scales. Using this characteristic, we can extract the medium-range order hidden beneath what otherwise appears as randomness.

The researchers built computational models of amorphous silicon by classical molecular dynamics wherein the temperature of the silicon was increased above the melting point and then gradually cooled (quenching) to room temperature. Differences in structural characteristics were introduced by changing the cooling rate.

Then, the persistent diagram, which is the two-dimensional visualization of persistent homology, was computed for each model. The researchers focused on that the diagrams reflect the structural features of amorphous silicon. Thus, they constructed the numerical representation, called ‘descriptors,’ that could be used in machine learning. The researcher found that the persistent diagram fulfilled the creation of a good descriptor for use in the machine learning procedure, which in turn achieved accurate predictions about the thermal conductivities.

By further analyzing the persistent homology data and machine-learning model, the researchers illustrated the previously hidden relationship between medium-range order in amorphous silicon and its thermal conductivity.

The study should now open an avenue for controlling material characteristics of amorphous silicon and other amorphous solids through the topology of their nanostructures.

Japanese researchers develop a model of motor learning that is able to simulate the results of experiments in humans

Tyler O'Neal, Staff Editor ECONOMICS

Even seemingly simple movements are very complex to perform, and the way we learn how to perform new movements remains unclear. Researchers from Japan have recently proposed a new model of motor learning that combines a number of different theories. A study published this month in Neural Networks revealed that their model can simulate motor learning in humans surprisingly well, paving the way for a greater understanding of how our brains work.

For even a relatively simple task, such as reaching out and picking up an object, there are a huge number of potential combinations of angles between your body and the different joints that are involved. The same goes for each of your muscles—there is an almost endless combination of muscles and forces that can be used together to act. With all of these possible combinations of joints and muscles—not to mention the underlying neuronal activity—how do we ever learn to make any movement? Researchers at the University of Tsukuba aimed to address this question.

The research team first created a mathematical model to imitate the learning process that occurs for new motor tasks. They designed the model to reflect many processes thought to occur in the brain when a new skill is learned. The researchers then tested their model by attempting to simulate the results of three recent studies that were conducted on humans, in which individuals were asked to perform completely new motor tasks.

"We were surprised at how well our simulations managed to reproduce many of the results of previous studies in humans," says Professor Jun Izawa, senior author of the study. "With our model, we were able to bridge the gap between a number of different proposed mechanisms of motor learning, such as motor exploration, redundancy solving, and error-based learning."

In their model, larger amounts of motor exploration—that is, variability in movements—were found to help with the learning of sensitivity derivatives, which measure how commands from the brain affect motor error. In this way, errors were transformed into motor corrections.

"Our success at simulating real results from human studies was encouraging," explains first author Lucas Rebelo Dal'Bello. "It suggests that our proposed learning mechanism might accurately reflect what occurs in the brain during motor learning."

The findings of this study, which indicate the importance of motor exploration in motor learning, provide insights into how motor learning might occur in the human brain. They also suggest that motor exploration should be encouraged when a new motor task is being learned; this may be helpful for motor rehabilitation after injury or disease.

Caltech scientists adapt methods from weather forecasting to assess the risk of COVID-19 exposure

Tyler O'Neal, Staff Editor ECONOMICS

A more granular understanding of risk could reduce the need for widespread lockdowns during an epidemic

Techniques used in weather forecasting can be repurposed to provide individuals with a personalized assessment of their risk of exposure to COVID-19 or other viruses, according to new research published by Caltech scientists.

The technique has the potential to be more effective and less intrusive than blanket lockdowns for combatting the spread of disease, says Tapio Schneider, the Theodore Y. Wu Professor of Environmental Science and Engineering; senior research scientist at JPL, which Caltech manages for NASA; and the lead author of a study on the new research that was published by PLOS Computational Biology  on June 23.

"For this pandemic, it may be too late," Schneider says, "but this is not going to be the last epidemic that we will face. This is useful for tracking other infectious diseases, too."

In principle, the idea is simple: Weather forecasting models ingest a lot of data—for example, measurements of wind speed and direction, temperature, and humidity from local weather stations, in addition to satellite data. They use the data to assess the current state of the atmosphere, forecast the weather evolution into the future, and then repeat the cycle by blending the forecast atmospheric state with new data. In the same way, disease risk assessment also harnesses various types of available data to assess an individual's risk of exposure to or infection with disease, forecasts the spread of disease across a network of human contacts using an epidemiological model, and then repeats the cycle by blending the forecast with new data. Such assessments might use the results of an institution's surveillance testing, data from wearable sensors, self-reported symptoms and close contacts as recorded by smartphones, and municipalities' disease-reporting dashboards.

The research presented in PLOS Computational Biology is proof of concept. However, its result would be a smartphone app that would provide an individual with a frequently updated numerical assessment (i.e., a percentage) that reflects their likelihood of having been exposed to or infected with a particular infectious disease agent, such as COVID-19.

Such an app would be similar to existing COVID-19 exposure notification apps but more sophisticated and effective in its use of data, Schneider and his colleagues say. Those apps provide a binary exposure assessment ("yes, you have been exposed," or, in the case of no exposure, radio silence); the new app described in the study would provide a more nuanced understanding of the continually changing risks of exposure and infection as individuals come close to others and as data about infections is propagated across a continually evolving contact network. 

The idea was born in the early days of the COVID-19 pandemic, when colleagues and partners Schneider and Chiara Daraio, the G. Bradford Jones Professor of Mechanical Engineering and Applied Physics and Heritage Medical Research Institute Investigator, abruptly found themselves isolating at home and wondering how to use their scientific and engineering expertise to help the world deal with this new threat.

One pre-pandemic focus of Daraio's research was the development of low-cost body temperature trackers. And that raised the question: Would the widespread use of such trackers allow for better tracking and understanding of COVID-19's spread?

"We were envisioning something like a weather forecasting app, harnessing information from sensors, infection data, and proximity tracking, which people could use to adjust their behavior to mitigate individual risks," says Daraio, co-author of the PLOS Computational Biology paper.

Schneider is a climate scientist who helms the Climate Modeling Alliance (CliMA), which is leveraging recent advances in the computational and data sciences to develop a wholly new climate model. He reached out to longtime acquaintance Jeffrey Shaman of Columbia University. Shaman's research on how climate change affects the spread of infectious diseases led Shaman to an interest in epidemiology and the adaptation of similar weather-forecasting methods for disease modeling on the community level.

"Over the last decade, the field of infectious-disease modeling, and forecasting in particular, has exploded. Many disease-forecasting approaches leverage ensemble and inference methods commonly used in weather prediction," says Shaman, co-author of the PLOS Computational Biology paper.

The team had two key challenges: adapting weather-prediction methods for this purpose and developing a realistic test bed to gauge how well it works.

"Conceptually it is a very appealing idea, as methods to forecast weather have been so effective in predicting the chaotic atmosphere, a famously challenging task," says Caltech research scientist Oliver Dunbar. "But there is no direct translation. An epidemic-forecasting app has very little data to work with and only on a partial population of users. We fortunately found success by coupling this sparse data with the latest smart-device technologies and a mathematical viral spreading model."

To test it, the team turned to Lucas Böttcher of the Frankfurt School of Finance and Management in Germany. Böttcher built a computer model of an imaginary city—a downscaled and idealized version of New York City—with 100,000 "nodes," or fictional people, and then studied how well the adapted weather-forecasting methods predicted the spread of a disease through the population.

The results were encouraging: in the simulations, the model identified up to twice as many potential exposures than would be caught by traditional contact tracing or exposure-notification apps when both use the same data. 

"The methods developed in our study are relevant not only in the context of infectious disease management, but they also open up new ways of combining observation data with high-dimensional mechanistic models arising in computational biology," says Böttcher, co-author of the PLOS Computational Biology paper.

Despite these promising results, the implementation of this technology in the real world requires suitable levels of smart-device users, and effective testing campaigns to make the risk-assessment software work for managing and controlling epidemics. If approximately 75 percent of a given population provide relevant information (for example, whether they have tested positive for a disease) and self-isolate when they may have been exposed, the risk-assessment software is accurate enough to manage and control the COVID epidemic throughout the entire population. And yet, as is evident by COVID-19 vaccination rates, buy-in by such a large fraction of the population is difficult to achieve. 

Nevertheless, a promising scenario is a deployment by smaller community user bases—for example, the population of a college campus—that can readily provide the software with more than enough data to provide accurate risk assessments that will locally reduce the spread of disease.

"The challenge in making this system a reality is managing privacy concerns, for example, about transferring data about close contacts to a central data-processing facility," Schneider says. "That being said, only anonymized information is needed. Location information is already routinely collected for commercial use, and we envision ways to harden the system against exploitation by bad actors."

Other co-authors of the PLOS Computational Biology paper include Caltech research scientist Jinlong Wu and graduate student Dmitry Burov as well as former Caltech postdoc Alfredo Garbuno-Iñigo of Instituto Tecnológico Autónomo de México; Gregory Wagner and Raffaele Ferrari of MIT (all members of CliMA); and Sen Pei of Columbia University. This research was supported by Eric and Wendy Schmidt and Schmidt Futures; the Swiss National Science Foundation; the National Institutes of Health; the Army Research Office; the National Science Foundation; the National Institute of Allergy and Infectious Diseases; and the Morris-Singer Foundation.