Machine learning
Uncertainty
2023 - On the Uncertainty Estimates of Equivariant-Neural-Network-Ensembles Interatomic Potentials
- Luca M. Ghiringhelli, https://arxiv.org/abs/2309.00195
2023 - Deep ensembles vs committees for uncertainty estimation in neural-network force fields: Comparison and application to active learning
- Jesús Carrete, J. Chem. Phys. 158, 204801 (2023)
- Rafael Gómez-Bombarelli, npj Computational Materials volume 9(225) 2023
2023 - Fast uncertainty estimates in deep learning interatomic potentials
- Boris Kozinsky, J. Chem. Phys. 158, 164111 (2023)
- Johannes Kästner, Digital Discovery, 2022, 1, 605
2020 - Methods for comparing uncertainty quantifications for material property predictions
- Zachary W Ulissi, Mach. Learn.: Sci. Technol. 1 (2020) 025006
- We need a powerful uncertainty measurement so that we know when to trust the predictions and how much we can trust on them.
Transfer learning
2021 - PIP delta learning from DFT to CCSD(T)
- Δ-machine learning for potential energy surfaces: A PIP approach to bring a DFT-based PES to CCSD(T) level of theory
- A. Nandi, C. Qu, P.L. Houston, R. Conte, and J.M. Bowman, J. Chem. Phys. 154, 051102 (2021).
- https://aip.scitation.org/doi/10.1063/5.0038301
The Δ-machine learning is represented as a equation: $V_{low \rightarrow high} = V_{low} + \Delta V_{high-low}$.
- $V_{low}$ is a fitted potential at low-level theory like DFT
- $\Delta V_{high-low}$ is a the energy difference between high-level theory like couple cluster and the low-level theory
The goal is to fit a high-level potential with a small amount of training data.
The direct fitting $V_{high}$ requires a lot of data and computational resources.
It would save computational costs during the accumulation of training data if one trains on $V_{low}$ on many low-level data and learns $\Delta V_{high-low}$ by a few high-level data.
The assumption here is that $\Delta V_{high-low}$ is less influenced by nuclear coordinations than $V_{low}$.