Show Phase I projectsShow Phase III projects

Inc 2 - Machine Learning

Supervised learning of materials properties of molecules and solid materials has been demonstrated to afford machine learning (ML) models which enable property predictions of new materials with low approximation error. Equally important, their computational cost is negligible when compared to conventional quantum methods, such as density functional theory (DFT). Over the first phase of MARVEL, von Lilienfeld and Ceriotti have made substantial progress on the development and application of unsupervised and supervised ML models. The findings have enabled the systematic and rigorous classification of materials, the identification and ranking of low-dimensional governing collective variables, as well as reaching an unprecedented combination of computational speed and predictive power across compositional and configurational compound space. The latter aspect subsequently enabled the discovery of ~90 crystals of the Elpasolite structure, predicted to be thermodynamically stable.

The project is led by Analole von Lilienfeld.

Group Leaders

Anatole von Lilienfeld
Project leader
UniBas, Basel
Volker Roth
Group leader
UniBas, Basel
Michele Ceriotti
Deputy director
EPFL, Lausanne

These observations suggest that ML methods can now be adapted to tackle the more generic challenge of materials discovery and design. In order to resolve the problem of extrapolation versus interpolation we plan to exploit that training and application of ML models is "agnostic" in the sense that origin and choice of the training data is irrelevant for formal construction of the ML model, due to its inherently inductive nature. As such, ML models can be improved or updated on a regular basis, through adaptation of training set size and composition, compound's representations, or locally adaptive regression functions. More specifically within this incubator project Inc 2 — Active Machine Learning for Computational Materials Design — we investigate possible ways to systematically accelerate the computational identification of promising materials candidates through combinations of supervised and unsupervised ML models, property optimization algorithms, and active learning. 

Related publications (until January 2023)