Modular development of deep potential for complex solid solutions

Quantum-Accurate Modeling of Ferroelectric Phase Transition in Perovskites from Message-Passing Neural Networks

“… a See refs , . b T C is from ref ; c / a and P S are from ref . c Estimated from Figure 1a in ref . d See Figure 2b in ref . e See Figure 5 in ref . f Estimated from Figure 2 in ref . g Estimated from Figure 3f in ref . h T C , c / a , and P S are from refs , and ref , respectively. …”

“…It is also important to compare MPNNs with the deep potential (DP) model. , The DP model is a deep neural network model similar to BPNNs but uses a local coordinate frame as its inputs instead of descriptors. There are several differences between the DP approach and the MPNN approach: (i) MPNNs use the graph methodology and learn the interaction between atoms through the message-passing scheme; (ii) In principle, the graph representation and the iterative message-passing process used in MPNNs make the atoms ultimately semilocal. , As the number of interaction blocks (specifically, the number of convolutional and aggregating operations) increases, the site energy of each atom is further influenced by the atoms beyond the cutoff distance.…”

“…Practically, we are interested in artificial neural networks (ANNs) that can be trained to efficiently predict interatomic potentials. Common ANN architectures for such purposes include the descriptor-based Behler–Parrinello neural networks (BPNNs), − the recently developed DeePMD, , and the graph-based message-passing neural networks (MPNNs). − We base our investigation on the MPNN methodology and construct a refined version of DimeNet++ (directional message-passing neural network), , a recent successful implementation of MPNNs, to model perovskites. We note that, in addition to ANNs, employing Gaussian process regression, the SOAP-GAP approach has also been applied to study the phase transition and thermodynamics of BaTiO 3 …”

Implementation and Validation of an OpenMM Plugin for the Deep Potential Representation of Potential Energy

“…MLPs are increasingly employed in a range of applications for material systems [87][88][89][90], yet their applications to biomolecular systems remain limited due to the absence of longrange interactions in MLPs. Consequently, the hybrid MLP/MM scheme emerges as a more practical approach for modeling biomolecules with MLPs.…”

Data Efficient and Stability Indicated Sampling for Developing Reactive Machine Learning Potential to Achieve Ultralong Simulation in Lithium-Metal Batteries

“…59 Wu and co-workers recently proposed an ModDP strategy which builds training data for two end points along a reaction with DP-GEN and then refining the training data of intermediate states using random configurations. 60 The philosophy is actually pretty similar to using bulk configurations in the interface, see Figure S6. So, our question Q2 is as follows: Is it helpful to use bulk configurations in the training set for studying interfaces?…”