In this article, we explore the impact of artificial intelligence (AI) on computational materials science and describe the evolution of Materials Design’s leading materials modeling environment, MedeA, towards an AI platform for Materials Science.
Spurred by the success of large language models such as ChatGPT, AI has taken the world by storm. With the ability of AI models to ingest, process, recall, and summarize vast quantities of knowledge in natural language, AI has now surpassed the capacity of any individual human to answer general knowledge-based questions. Humanity has effectively pooled much of its collective knowledge, creating what could be called superhuman intelligent capabilities. Based on its vast knowledge, this type of AI can even be creative, effortlessly generating prose, poetry, computer code, images, or music based on just a few prompts.
But can AI do science?
Despite the remarkable achievements of AI, many aspects of scientific endeavor remain where AI is a useful but imperfect tool, and human scientists and engineers are far from being outsmarted just yet. This is largely because AI requires large amounts of data to train reliable models, and such data is not usually available in previously unexplored fields that are the domain of scientific discovery. While scientific discovery often involves mining existing knowledge, it is the combination of this knowledge with intuition, rigorous planning of experiments, and luck that typically leads to new discoveries. AI excels at mining knowledge and analyzing data, but humans still master the art of discovering the unknown.
One way scientists have traditionally encoded existing knowledge to aid new discoveries is through “computer modeling.” In various disciplines—from climate science and meteorology to chemistry, medicine, biology, engineering, and materials science—virtual models (recently termed “digital twins”) of the real world have become indispensable, complementing or replacing real experiments. Fundamentally, computer modeling involves taking physical equations—such as those governing quantum mechanics, fluid dynamics, reaction kinetics, or finite element deformations—and allowing a computer to solve these equations for specific systems.
This digital twin approach has proven tremendously successful. In materials research, it allows the prediction of properties and simulation of material behavior based solely on the atomistic structure of a material and the laws of physics (“ab initio modeling”). Established technologies of choice include Density Functional Theory and Molecular Dynamics. Materials Design’s MedeA Environment has long provided an easy-to-use interface to set up and run such calculations on the latest computer hardware, utilizing industry-leading implementations of these technologies.
Due to the complexity of real systems and the resulting computational cost of “ab initio” computations, scientists have added a second set of tools to their digital twin toolbox: empirical models describing statistical correlations. For example, the “Quantitative Structure-Property Relationships” (QSPR) method uses machine-learned models that describe materials properties as functions of structural “descriptors” or “fingerprints.” Model parameters relating properties to these descriptors are often derived by training the model against experimental data. While this approach may not trace materials properties back to their quantum-mechanical origins, it often aids remarkably well in discovering materials or molecules with desirable properties. Examples of MedeA technology for this type of empirical modeling include the MedeA QT (for QSPR-type models) and the MedeA Polymer Expert (for de novo polymer design).
The real promise of recent AI technological advances in materials science lies in the combination of these two approaches: “ab initio” and empirical machine learning (ML). The ability to run large numbers of first-principles calculations on the latest generation of supercomputing hardware has helped overcome the lack of experimental data that has previously stymied the development of empirical ML models in materials science. A race is now on to develop accurate and fast ML models of materials that are no longer based on fundamental physical equations. Instead, they utilize neural network models like those used in large language models to predict properties or describe atomistic interactions. These models are often trained against ab initio computations instead of experimental data. One such approach involves using interatomic machine-learned potentials (MLP) to simulate atomistic processes over much larger time and length scales than what is feasible with an “ab initio” approach. Once trained for a specific set of systems, these models can describe physical reality extremely well at a fraction of the cost of ab initio calculations. The MedeA Environment offers industry-leading capabilities for training and utilizing such MLPs.
Hence, while AI may not quite do science yet, it can certainly accelerate scientific discovery tremendously.
However, one barrier to adoption has been the sophistication and deep knowledge required to train AI models and provide the right parameters, systems, and file formats. What is needed is an AI platform that removes the need for advanced data science knowledge in developing and applying this powerful technology.
This is where Materials Design’s MedeA Environment excels. MedeA guides users through the entire machine-learning process within a single easy-to-use modeling and AI platform. Building on its pedigree as the leading environment for traditional materials modeling, MedeA now offers well-validated and intuitive interfaces and models to perform the training of MLPs and other ML models, featuring some of the most sophisticated ML architectures on the backend. This powerful combination enables scientists and engineers to easily conduct digital experiments on new materials.
In summary, the MedeA materials modeling and AI platform allows users to:
Easily build digital twins of atomistic systems.
Access the latest iterations of impactful computer modeling algorithms developed since the late 20th century, such as density functional theory and molecular dynamics (VASP, LAMMPS, and derived property calculators).
Easily train and publish ML models that relate atomistic descriptors to industrially relevant properties.
Utilize the latest HPC hardware to perform similar experiments on thousands of systems in parallel, for example, to screen materials for energy storage or desired mechanical or electronic properties.
Access pretrained models for property predictions. The MedeA Environment can even propose materials based on specified desired properties, as in Polymer Expert.
Easily train and use MLPs based on ab initio calculations and employ these MLPs to run large-scale computations of systems previously inaccessible to traditional computer modeling.
So can AI do science? Many would argue that while AI can (almost) drive your car, it still takes a human to imagine and develop a shoebox-sized battery to power that car for 1,000 miles on a single charge or to create a room-temperature superconductor-based levitating car seat for a more comfortable ride. But AI and ML will play a major role in such discoveries. Materials Design’s leading materials modeling and AI platform puts this technology at your fingertips.
Comments