The first full-fledged version of multifidelity machine learning software has been released by Vivin Vinod under the MIT license at PyPi for use in quantum chemistry applications. With a detailed documentation of the library and examples at https://vivinvinod.github.io/mfml-4-qc , the lightweight NUMBA accelerated package provides users with easy to deploy resources of multifidelity machine learning. Flexibility of machine learning architecture and cutting edge integration of numerical compute engines such as ORCA make mfml-4-qc a highly modular package. Active learning using the efficient Low Fidelity as Bias technique is also provided.
Author: Matthias Menge
Chemical space sampling with novel active learning cuts training data cost by an order of magnitude
Vivin Vinod and Peter Zaspel have developed a faster, more efficient way to train machine learning models on complex chemical data. Detailed in their new paper, “LFaB: Low Fidelity as Bias for Active Learning in the Chemical Configuration Space” in the Journal of Chemical Theory and Computation, they introduced a novel bias-based active learning strategy. By using a lower-fidelity output as a proxy for model bias, their method dramatically outperforms both traditional variance-based active learning and standard random sampling. The LFaB method is highly adaptable, proving its effectiveness across a wide range of chemical properties including excitation energies.