## Machine learning potential Current issue: DFT calculation is computationally expensive, ideal to use classical potential for larger systems. But even for traditional potential: - different systems require different potentials, like EAM for metal, some other for strong covalent material and for chemical reactions. - Training of potential required (fitting data) - Hard for multicomponent system - Have to recompute potential for new atoms/elements added - For potential has good chemical/physical natures, the transferability would be good. - Interpolation typically better than extrapolation. >[!Notice] >It's worth show the classification of interatomic potential. >- general-purpose potential >- specialty-purpose >- artificial potential (by property-based training) >>not directly from chem/phys, used to generate/evaluate specific phenomenon/property So people introduce machine learning potential (ML potential): - Map 3N-dimensional configuration onto its [[Potential energy surface]] (PES) - Optimize the regression parameters to obtain smooth potential energy surface that best interpolates between reference energies. $R_{i} \rightarrow G_{i} \stackrel{\text{R}}{\rightarrow} E_{i}$ Here $R_{i}$ is variable-size positions, and the first arrow is like [[Smooth overlap of atomic positions|SOAP]]; $G_{i}$ is the structural descriptors, aim to maintain the invariance and smoothness of PES. Then after the regression model represented by the second arrow, we have the energy. So the whole process is like: atoms (data set) > local structural parameters > regression > PES. ML potential does not have physics basis, so it is accurate, widely applicable, but poor transferability (poor extrapolation).