News

When we train Machine Learning models on quantum chemical data, we face a trade-off between generating large amounts of training data with approximate (e.g. semiempirical) methods or smaller amounts of higher quality data with expensive first-principles methods. This is problematic because both the amount and quality of the data ultimately determine how useful a ML model can be. Mengnan's new paper in MLST shows that we can overcome this dilemma to some extent by combining data of different fidelity levels in a transfer learning approach. This reveals that lower fidelity data can be useful to improve the performance of a model when only few high-fidelity data points are available.