The application of Artificial Intelligence (AI) to the scientific domain is expected to have transformative effects across numerous research areas. In particular, Condensed Matter Physics and Materials Science are among the research fields that expectedly will benefit the most from the adoption of new AI techniques. AI as a research topic or as a tool is starting to be adopted by many research groups, but is still not generalized despite its potential to be applied to a variety of problems.
From IFIMAC, the Spanish Network of AI for Condensed Matter Physics and Materials Science was launched in January 2020. It is an ongoing initiative coordinated by Jorge Bravo-Abad, IFIMAC researcher and member of the Universidad Autónoma de Madrid (UAM) Theoretical Condensed Matter Physics Department. The network aims at boosting the adoption of AI in Condensed Matter Physics and Materials Science in Spain. The network arises as an initiative to unite all efforts to apply the power of AI to the field of condensed matter physics and materials science.
The Spanish Network of AI for Condensed Matter Physics and Materials Science focuses on three main lines of activity. Firstly, the creation of an on-line platform that gathers and connects Spanish research groups and companies working on the application and development of AI techniques to scientific problems, particularly focusing on Condensed Matter Physics and Materials Science.
Next, the network aims to develop and maintain an open repository in which the members of the network can share algorithms, codes and databases that can help in the application of AI to scientific problems. Finally, efforts in advocacy and dissemination regarding the use of AI in Condensed Matter Physics and Material Science are to be undertaken, via seminars, training sessions, summer schools and specific outreach activities.
AI: a matter not only about materials science
In the field of AI, new tools are created, and new cross-discipline frameworks are created in which physicists, chemists, ecologists, economists and others are using the same AI tools (perhaps with field-specific particularities). This diversity is also present in materials science itself, also a focus point of the network launched by IFIMAC: quantum computing, material hardness, engineering materials, etc. all work with and benefit from these approaches.
AI skills are increasingly becoming a lever to position professionals from "pure sciences" in the labour market. “They are familiar with the language used in the digital industry. An important aspect is the transition of the language used between people from different areas. The common ground, would be AI and its concepts.” says researcher Jorge Bravo. “Such skills are a lever for career advancement in various industries.”
The takeoff of AI, continues Bravo, is almost exclusively due to the development of techniques of deep learning, which can be classified in three groups: supervised learning, which entails learning from labeled data; unsupervised learning, which uses unlabeled data and reinforcement learning, which learns based on feedback from the environment.
Programming, data processing, math and knowledge of the tool ecosystem are foundations which combine with the particular domain knowledge of the scientist, who subsequently applies this to research problems. “In science you often are constrained into a context of working with a small amount of data. AI, integrating specific domain knowledge, can in fact help to operate with few data. This works by integrating the previous insights of a given research field taking them into account in algorithms”, continues Bravo. Such interaction of previous knowledge with AI approaches unlocks improved accuracy, stronger prediction power and potentially the ability to complete far more work in less time.
About AI and the “blackbox problem”
A recurrently discussed problem in the area of AI is the unveiling of how in effect machines “learn”, and what rules they incorporate on basis of their training data. Currently AI algorithms learn as "black boxes" in which it is impossible to know in any detail, even at the mathematical level, what is going on during a deep learning process. This lack of control is problematic as it may yield unexpected and even negatively disruptive results when operating outside the planned work settings.
As an example, a single neuron in a neuronal network cannot in effect learn anything; however, many neuron units learning together can do so, in a collective learning phenomenon that unfortunately cannot currently be tracked in detail. Statistic physics, a branch of condensed matter physics, studies systems in which many units of a given element conform together a new macroscopic entity with emergent properties derived from the collective of individuals.
“Statistic physics is starting to become a tool to understand the deep learning process. This is an unexpected synergy of condensed matter physics with the area of AI as a whole and which may help solve the unknowns of the learning processes of AI algorithms.” The expertise in condensed matter physics could hence play a pivotal role in helping to understand the AI learning process better.
“It is interesting how AI is penetrating manifold discilines in a short period of time, and this it is a very transversal phenomenon. The same language and concepts start to be used by biologists, materials scientists or engineers, which can be an amazing unifying force to facilitate interdisciplinary research. The tools need to be studied, the mindset needs to be picked up as well, and this is surely happening yet at most research centres of excellence. This may open the door to new paradigms of how research is undertaken... this is very unique”, finishes Bravo.
Banner of the Spanish Network of AI for Condensed Matter Physics and Materials Science reproduced with permission of IFIMAC.
AI diagram kindly provided by Jorge Bravo.
Neural network concept diagram was downloaded from Flickr and licensed via a Creative Commons Attribution 2.0 Generic (CC BY 2.0) license.