.png){kind=icon}
Return
Lifelong Machine Learning Potentials
.png){kind=icon}
OA
AI
DOI：10.1021/acs.jctc.3c00279.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Machine learningpotentials (MLPs) trained on accurate quantumchemical data can retain the high accuracy, while inflicting littlecomputational demands. On the downside, they need to be trained foreach individual system. In recent years, a vast number of MLPs havebeen trained from scratch because learning additional data typicallyrequires retraining on all data to not forget previously acquiredknowledge. Additionally, most common structural descriptors of MLPscannot represent efficiently a large number of different chemicalelements. In this work, we tackle these problems by introducing element-embracingatom-centered symmetry functions (eeACSFs), which combine structuralproperties and element information from the periodic table. TheseeeACSFs are key for our development of a lifelong machine learningpotential (lMLP). Uncertainty quantification can be exploited to transgressa fixed, pretrained MLP to arrive at a continuously adapting lMLP,because a predefined level of accuracy can be ensured. To extend theapplicability of an lMLP to new systems, we apply continual learningstrategies to enable autonomous and on-the-fly training on a continuousstream of new data. For the training of deep neural networks, we proposethe continual resilient (CoRe) optimizer and incremental learningstrategies relying on rehearsal of data, regularization of parameters,and the architecture of the model.
Keywords:
FORCE-FIELD
NEURAL-NETWORKS
UNCERTAINTY
ACCURACY
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
5.5
Papers:
1.1W
Citations:
5.4W
Organization
Cited Papers
ShareSaveSelf-assembly of five new organic–inorganic hybrids based on two new flexible tricationic templates
PREAI
ShareSave
ShareSaveSpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects.png)
NATURE COMMUNICATIONSOAAI
NATURE COMMUNICATIONSIF15.7
OAAIShareSaveMachine learning enables long time scale molecular photodynamics simulationsChemical ScienceOAAI
Chemical ScienceIF7.4
OAAIShareSaveApproaching coupled cluster accuracy with a general-purpose neural network potential through transfer learningNATURE COMMUNICATIONSOAAI
NATURE COMMUNICATIONSIF15.7
OAAIShareSave
View more