Shock temperature fields no longer tied to one strength: MISTnet2 unifies atomistic and coarse-grain shock simulations with a conditional U-Net
Synopsis
MISTnet2 extends prediction of the three-dimensional temperature field in shocked energetic materials from a single shock strength and a single simulation method to a continuous range of piston velocities and two simulation fidelities. The earlier MISTnet handled one piston velocity and erred by order 1000 K on hotspots from pores in RDX crystals. The new model shares one conditional 3D U-Net, conditioned on simulation type, piston velocity and voxel size, and matches simulations within its training distribution while giving only qualitative trends outside it.
Interpretation
Machine-learned prediction of hotspot temperature fields in shocked energetic materials now covers atomistic molecular dynamics and coarse-grain simulation data in one network, across a continuous range of shock strengths. The earlier MISTnet corresponded to a single shock strength, a single simulation method and one class of microstructure, and predicted the hotspot of a single pore in RDX only qualitatively, with maximum-temperature errors of order 1000 K. For atomistic polymer-bonded explosive and coarse-grain porous RDX, cumulative temperature-versus-volume curves agree with simulations on both training and test sets, and the model identifies all hotspots, including the smallest and hottest and the large moderate-temperature ones.
The model maps an initial microstructure directly to the post-shock temperature field, taking binned RDX density, polystyrene density and their product as input and outputting temperature in kelvin divided by 1000. Conditional inputs are piston velocity, simulation type and voxel size, with simulation type encoded as a two-element one-hot vector separating atomistic molecular dynamics from dissipative particle dynamics. Atomistic data come from RDX/polystyrene slabs of 61 nm by 18 nm by 100 nm at about 1.49 g/cm3 with piston velocities of 1.0, 1.5 and 2.0 km/s; coarse-grain data are pure RDX slabs of 500 nm by 4 nm by 1 micrometer containing pores of 100, 50 and 20 nm diameter.
Within the training distribution the model interpolates to unseen shock strengths, while at higher shock strength it underestimates the temperature rise. Interpolation at 1.25 and 1.75 km/s is highly accurate, whereas extrapolation to 2.25 km/s correctly predicts higher temperatures but underestimates the effect. All of these evaluations use test data excluded from training, and the authors caution against quantitative use outside the training distribution, calling for additional data at higher piston velocities or pores larger than 100 nm.
Perspective
The demonstrated scope is RDX-based energetic materials: atomistic RDX/polystyrene polymer-bonded explosive and coarse-grain porous RDX crystals, piston velocities of 1 to 2 km/s, pore diameters of 20 to 100 nm, with temperature fields represented on a binned grid. The authors propose expanding the training data to improve extrapolation and adding material quantities such as shock-induced pressure and strain.
The authors explicitly caution that the model is unreliable outside the training distribution and gives only qualitative trends there; voxel-level deviations on the coarse-grain test set exceed those on the training set, which the authors attribute to jetting, convection and plastic deformation around large pores and sharp interfaces; near the 2 km/s criticality boundary, three models trained with different seeds predict deflagration in 3, 3 and 4 microstructures, and cases disagreeing with simulation come into agreement when the piston velocity changes by 0.1 km/s.
