The project has already achieved significant milestones, including addressing fundamental mathematical issues in atomic structure descriptions, developing robust long-range models, and achieving substantial progress in integrating ML with QM and SM methods (ML1). We already started releasing some of the planned software tools, making these advancements accessible to the broader scientific community (SW).
Simulations of transport in solid-state electrolyte materials and of the thermodynamics and dielectric behavior of prototypical ferroelectrics provide early proof of the relevance of this line of research to develop technological solutions that are of great relevance for contemporary societal issues (APP).
One result worth highlighting is the derivation of a method to achieve a complete local description of an atomic structure based on finite-correlation-order descriptors.
Soon after submitting the proposal for this project, we realized that one of the basic assumptions we made, the widely shared belief that the *local* descriptors used in the field were a solved problem, did not hold. We found a few arrangements of atoms that could not be distinguished by those descriptors - meaning that the ML models built on them were not able, even in principle, to provide a universal approximation of the quantum mechanical structure-property relations. Fully understanding -- and resolving -- this problem was one of the early achievements in FIAMMA, which allowed us to stay on track with the original plans.
With this potential roadblock removed, we could make excellent progress on the three methodological work packages: We've developed models that accurately describe different types long-range interactions (ML1), and integrated machine learning Hamiltonians into quantum mechanical calculations (ML2). Our work the integration of ML models with sophisticated statistical mechanical calculations has also yielded promising early results (ML3), which allowed us to secure additional funding to go beyond what initially envisaged to tackle the issue of assessing the level of trust that can be put in ML-driven predictions in a reliable and inexpensive way.
On the software-development front, we have completely overhauled our ML software infrastructure, making it more modular and efficient. Most of the core elements are still in heavy-development, but already openly accessible. We have already released some production libraries, e.g. Sphericart for efficient spherical harmonics computation and Scikit-Matter for domain-agnostic machine learning methods. We have also finalized a new release of the i-PI advanced simulation code, focusing on ML model integration (SW, ML3).
In terms of applications (APP), we've slightly deviated from our original plans in terms of the target systems, but have already made impactful discoveries.
Our focus shifted to ferroelectrics and lithium thiophosphate electrolytes, materials that allowed us to explore long-range physics and electronic properties while establishing valuable collaborations with experimentalists and industry partners. Although we intend to return to our initial focus on silicate frameworks and molecular materials, our current applications have effectively demonstrated the practical implications of our methodological advancements.
