An Extreme-Scale Multi-Fidelity Computational Active Learning Paradigm Towards Realizing Autonomous Synthesis

With the advent of automated experiments, controlling synthesis to create new materials has become a key challenge for the materials community. This is a non-trivial problem due to lack of governing equations that can describe syntheiss of materials. While theoretical simulations have largely been used to discover materials with targeted properties, we hypothesize that simulated synthesis can be used to eventually guide real experimental synthesis. This requires relating thermodynamical and kinetic drivers (e.g., temperature, pressure, chemical potential etc) of material synthesis to synthesized material structure and proeprties. This tedious and time-consuming approach becomes particularly cumbersome to study simulated synthesis of chemicals and materials with complex dependencies on local environment, temperature and lattice-strains e.g., heterostructures or interfaces of nanomaterials. In this talk, I will present workflows that couple automated exascale high-throughput large-scale DFT calculations, ensemble force-field fitting and molecular dynamics simulations with a wide range of uncertainty quantification-driven active learning paradigms for on-the-fly learning of material synthesis trajectories, to create an autonomous computational synthesis platform. By implementing such a workflow to study recrystallization of amorphous transition-metal dichalcogenide (TMDC) phases under various growth parameters, I will show that such automated multi-fidelity frameworks can be promising towards achieving controlled epitaxy of targeted multilayer moiré devices paving the way towards a robust autonoumous discovery pipeline to enable unprecedented functionalities. Opportunities to use these autonomous computational synthesis pipelines to create 'digital twins' of synthesis trajectories, train generative inverse-design machine-learning algorithms to combine material discovery with targeted proeprties and their synthesis in a single framework and eventually accelarate experimental synthesis will also be presented.