Blending Artificial Intelligence and Physical Sciences: What Can We Expect?

From virtual assistants like Siri to self-driving cars and computer-aided medical diagnoses, artificial intelligence (AI) is affecting our lives with unprecedented speed. Slowly but steadily, scientists in a broad range of fields have started to embrace AI in their research, hoping to significantly reduce the time needed to achieve new discoveries. This trend has become more obvious in the physical sciences, and in the field of materials science in particular, which is focused on the discovery and production of new, advanced materials imbued with desirable properties or functions. Think: screens of foldable smartphones; batteries that power electric cars; or materials that bend light around them, rendering them invisible.

How exactly could AI help materials scientists? We recently interviewed three honorees of the Blavatnik Awards for Young Scientists, Alexandra Boltasseva, PhD, Professor of Electrical and Computer Engineering at Purdue University; Léon Bottou, PhD, Principal Researcher at Facebook AI Research; and Sergei V. Kalinin, PhD, Corporate Fellow at Oak Ridge National Laboratory, who are contributing to an upcoming virtual symposium on October 6 and 7, AI for Materials: From Discovery to Production. Here’s what they had to say about the opportunities, as well as the challenges, in this rising field.

It is only recently that researchers in the physical sciences, like materials scientists, have begun to incorporate AI techniques into their work. Why do we need to take advantage of AI for this field? What benefits may AI offer materials science?    

Sergei: AI offers a set of powerful tools to explore large volumes of multidimensional data in the physical sciences, and promises to uncover hidden functional relationships between the physical properties that we can observe. As such, AI methods are poised to become an inseparable part of all physical sciences, to enable discovery and hypothesis-driven research and to guide planning of experiments. We can take advantage of a broad range of AI techniques—from multivariate statistics to convolutional networks, unsupervised and semi-supervised methods, Gaussian processing, and reinforcement learning.

In addition, the proliferation of laboratory automation in areas from materials synthesis to imaging of materials’ molecular structures opens up broad opportunities for AI-driven experiments. For example, we will be able to adopt large-scale robotic systems or the microscale lab-on-a-chip platforms in our experiments, producing thousands or more materials in a single process. 

Alexandra: My own field, photonics, has truly been transformed by the concept of “inverse design,” meaning scientists input desired performances of photonic systems into computers and run physics-informed algorithms to figure out the best possible optical designs. The daunting challenge of this field lies in the inconceivably high computational power required for an exhaustive search within the extremely large, hyper-dimensional space of optical design parameters and constituent materials. Merging AI techniques with photonics is expected to not only enhance and enrich the design space, but, most importantly, to unlock novel functionalities and bring about disruptive performance improvements.

As compared to life sciences and pharmaceutical sciences, the application of AI in physical sciences is at least 10 years behind. What do you think is the biggest challenge for applying AI in physical sciences? How could the AI and physical sciences communities work together to address these challenges? 

Léon: Using machine learning in physical sciences is not an obvious proposition. Recent advances in AI have shown how tasks in computer science, such as computer vision and machine translation, can be achieved using big data. Yet it would be unwise to claim that this success can be replicated in all scientific fields. Big data only reveal statistical correlations that are not always indicative of the causal relations that physicists often seek. To solve this question, the AI and physics communities may take the strategy of defining a hierarchy of problems for which one could envision using AI, such as:

• Visualizing or measuring an ongoing physical phenomenon. These problems are the most accessible to AI/machine learning because they can directly leverage recent advances in computer vision and signal analysis in collecting data from physical experiments and computations.

• Explaining a physical phenomenon. These problems belong to the next rung of difficulty because we need AI/machine learning systems that incorporate enough of our current knowledge of physics, and can then clarify the phenomenon of interest by constructing something interpretable on top of our current knowledge.

• Designing a physical system that leverages a certain phenomenon in new ways. These are by far the most difficult problems, because they require AI/machine learning systems to accurately predict how the physical phenomenon will be affected by changes that are not included or prominent in the experimental data on which AI models have been trained.
  

Alexandra: The physical sciences community should ultimately build extensive databases to unleash the power of AI. We should even set up an ‘optical structures and materials genome’ project to construct a comprehensive dataset of photonic concepts, architectures, components and photonic materials to enable hierarchical machine learning algorithms that could provide ultimate-efficiency devices.

Sergei: I agree with Alexandra. AI tends to proliferate in the communities that adopt the model of open sharing of codes and data. While some areas of physics research have undergone this transformation, many more require both enabling tools and proof-of-benefit to accelerate this process.

I also want to add on to Léon’s comment on the fundamental difference between the AI and physics communities. AI starts with purely correlative models, and tends to rely on big data. In comparison, research in physical sciences is strongly based on prior knowledge to explore the cause and effect relationships, and often assumes the presence of simple rules or descriptors that can give rise to complex behaviors in macroscopic systems. Experiments in physical sciences can give rise to huge data volumes, but these data can pertain only to one specific situation of the system and hence are not “big.”

In order to further leverage the benefits of AI in physical sciences, researchers have to possess both sufficient domain knowledge in physical sciences and expertise in machine learning, or forge robust interdisciplinary collaborations. Conferences like AI for Materials will help researchers in both fields form these kinds of interdisciplinary teams.