Specialising neural network potentials for accurate properties and application to the mechanical response of titanium

Abstract

Large scale atomistic simulations provide direct access to important materials phenomena not easily accessible to experiments or quantum mechanics-based calculation approaches. Accurate and efficient interatomic potentials are the key enabler, but their development remains a challenge for complex materials and/or complex phenomena. Machine learning potentials, such as the Deep Potential (DP) approach, provide robust means to produce general purpose interatomic potentials. Here, we provide a methodology for specialising machine learning potentials for high fidelity simulations of complex phenomena, where general potentials do not suffice. As an example, we specialise a general purpose DP method to describe the mechanical response of two allotropes of titanium (in addition to other defect, thermodynamic and structural properties). The resulting DP correctly captures the structures, energies, elastic constants and γ-lines of Ti in both the HCP and BCC structures, as well as properties such as dislocation core structures, vacancy formation energies, phase transition temperatures, and thermal expansion. The DP thus enables direct atomistic modelling of plastic and fracture behaviour of Ti. The approach to specialising DP interatomic potential, DPspecX, for accurate reproduction of properties of interest “X”, is general and extensible to other systems and properties.

Deep Potentials for Materials Science

“…The screw a dislocation is more stable on Pyramidal I plane than on Prism plane for DP, which agrees with previous DFT calculations and experimental measurements [120]. We refer readers to [103] for more details. This example demonstrates the importance of specialisation of the DP and the superior flexibility of DP over other empirical potentials.…”

“…[σ lo ,σ hi ] Mg [78] [0.03,0.13] Al [78] & Al-Mg [78] [0.05,0.15] Cu [79] [0.05,0.20] Mg-Al-Cu [102] [0.05,0.20] Ti [103] [0.10,0.25] at T < 1.5Tm a for bulk exploration and [0.15,0.30] elsewhere W [104] [0.20,0.35] Ag-Au [105] [0.05,0.20] water [96] [0. ) and the number of starting structures should also be small (e.g., 5 each for different distorted crystal supercells).…”

“…For some particular applications, the DP must be tuned or specialised for important, yet subtle properties. This may be accomplished through specialisation, as discussed below [103].…”

“…Figures 7(a) and (b) display the speed comparison of the compressed Ti DP [103] with an EAM [115], and an MEAM potential [116] on a CPU and GPU machine. Note that the DP has a larger radius cutoff distance than EAM and MEAM in this case.…”

Benchmarking Structural Evolution Methods for Training of Machine Learned Interatomic Potentials

“…In the MLIP literature, it is common for MLIPs to perform well in reproducing reaction pathways for diffusion and stacking fault migration in metallic systems, but less so for insulating or molecular systems. [33][34][35][36][37][38] Some authors have taken to more advance ML techniques to address the problems of reaction pathways in molecular systems such as the recent work by Sun et al 39 These short-comings will need to be addressed before MLIPs can be reliably used to treat transition state phenomena, which motivates exploring other methods such as active and adversarial learning in the future.…”