Microstructure-Aware Deep Learning Bridges Atomistics to Macroscale for Shock-to-Detonation Prediction

Aidan Pantoya; Alejandro Strachan; Chunyu Li; Marisol Koslowski; Simon Gonzalez-Zapata

arxiv: 2605.27325 · v1 · pith:MRV2VZ43new · submitted 2026-05-26 · ❄️ cond-mat.mtrl-sci

Microstructure-Aware Deep Learning Bridges Atomistics to Macroscale for Shock-to-Detonation Prediction

Simon Gonzalez-Zapata , Aidan Pantoya , Chunyu Li , Marisol Koslowski , Alejandro Strachan This is my paper

Pith reviewed 2026-06-29 16:33 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords deep learningmolecular dynamicsfinite elementshock-to-detonationenergetic materialsRDXmicrostructuremultiscale modeling

0 comments

The pith

A convolutional neural network trained on molecular dynamics supplies sub-grid shock physics to finite-element models for parameter-free detonation prediction in RDX composites.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MISTnetX, a convolutional deep neural network trained on molecular dynamics simulations of shock waves traveling through complex microstructures. The network learns to recognize shock-microstructure interactions, hotspot formation, and the onset of deflagration, then supplies this information as sub-grid input to larger-scale finite-element simulations that track mechanics, thermal transport, and chemistry. This direct link addresses the absence of scale separation in shock-to-detonation problems, where processes range from angstroms to millimeters. The method is shown on a synthetic nanostructured plastic-bonded RDX composite, where it produces a complete run-to-detonation prediction without any fitted parameters.

Core claim

MISTnetX captures shock-microstructure interactions, hotspot formation, and the transition to deflagration from MD simulations and supplies this critical sub-grid information to FE simulations of mechanics, shocks, thermal transport, and chemistry. Applied to a synthetic but realistic nanostructured plastic-bonded RDX composite, MISTnetX enables parameter-free prediction of the full run-to-detonation transition.

What carries the argument

MISTnetX, a convolutional deep neural network trained on MD simulations of shock propagation through microstructures, which extracts and supplies sub-grid hotspot and transition data to continuum FE models.

If this is right

The coupled MD-FE workflow predicts the entire shock-to-detonation process without adjustable parameters once the network is trained.
Microstructure details from the atomistic scale directly influence macroscopic detonation outcomes in the model.
The same network architecture can supply information on mechanics, thermal transport, and chemistry within the continuum solver.
The approach works for realistic but synthetic nanostructured energetic material geometries.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the method holds, detonation sensitivity could be predicted from measured or imaged microstructures rather than from bulk empirical data.
Similar network bridging could be tested on other materials where atomistic details control macroscopic failure, such as fracture in composites.
Direct comparison of predicted hotspot statistics against high-resolution imaging during shock loading would test the network's internal representations.

Load-bearing premise

The molecular dynamics simulations accurately capture all relevant shock-microstructure interactions and the trained network generalizes without artifacts when used inside the finite-element simulations.

What would settle it

An experimental measurement of run-to-detonation distance or time in the same nanostructured plastic-bonded RDX composite that differs substantially from the MISTnetX-coupled FE prediction would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.27325 by Aidan Pantoya, Alejandro Strachan, Chunyu Li, Marisol Koslowski, Simon Gonzalez-Zapata.

**Figure 2.** Figure 2: Evolution of a STDT for an initial impact of 𝒖𝒑 = 𝟏. 𝟕𝟓 km/s. Key steps of ignition, growth and coalescence of deflagration waves, and the launch of a secondary shock are clearly visible before the transition to detonation. The time evolution of pressure profiles, see Fig. 3a for 𝑢" =1.75 km/s results, provides a clear picture of the STDT. The solid black lines represent the average pressure of the cross s… view at source ↗

**Figure 3.** Figure 3: Shock to detonation profile. (A) Pressure-position profiles at different times for an initial impact velocity of 1.75 km/s. Subsequent profiles are spaced by 2.5 ns, with the initial one corresponding to t = 0.25 ns after the initial impact. The last two lines, t = 35 and 30 ns, show the stead detonation pressure profile. (B) Temperature-position profiles at the same time as (A). The temperature profiles s… view at source ↗

**Figure 4.** Figure 4: 2D and 3D visualization of the shock to detonation transition. (A) Temperature, reaction extent, and pressure field during the early stages of the STDT showing hotspot ignition and growth. (B) Temperature, pressure, and microstructure 3D profiles showing DDT at later stages of STD transition. (C) Reconstructed Hugoniot from a 1.75 km/s shock. Discussion A novel approach to learn microstructural effects fro… view at source ↗

**Figure 5.** Figure 5: Pressure over different SDT stages for a shocked multi-fidelity PBX microstructure. Three stages are depicted, starting with ignition, then deflagration, then detonation. Our approach can be generalized to other problems in multiscale modeling. Examples include the mechanical response of advanced alloys (38) and ultrasonic initiation of chemistry (39). Finally, we note that the underlying model does not ne… view at source ↗

read the original abstract

The shock-to-detonation transition in energetic materials is governed by coupled processes spanning Angstroms to millimeters and femtoseconds to microseconds, where traditional multiscale models fail due to the lack of scale separation. We address this grand challenge by directly bridging large-scale molecular dynamics (MD) simulations with continuum finite-element (FE) models using MISTnetX, a convolutional deep neural network. Trained on MD simulations of shock propagation through complex microstructures, MISTnetX captures shock-microstructure interactions, hotspot formation, and the transition to deflagration, supplying critical sub-grid information to FE simulations of mechanics, shocks, thermal transport, and chemistry. Applied to a synthetic but realistic nanostructured plastic-bonded RDX composite, MISTnetX enables parameter-free prediction of the full run-to-detonation transition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MISTnetX trains a conv net on MD shock data to supply sub-grid terms to FE models for the full detonation transition, but the parameter-free claim hinges on unverified MD accuracy and network generalization.

read the letter

The main takeaway is that the authors train a convolutional network called MISTnetX on molecular dynamics runs of shocked microstructures and use it to feed sub-grid information on hotspots and chemistry into a finite-element code. This lets the continuum model run the entire shock-to-detonation process in a synthetic plastic-bonded RDX composite without the usual scale-separation assumptions.

The approach is new in its direct application to the detonation transition problem. Prior surrogate networks have been used in other multiscale settings, but the specific coupling here for mechanics, thermal transport, and reaction progress in energetic materials is a distinct step. The paper correctly identifies the core difficulty: processes that span angstroms to millimeters and femtoseconds to microseconds.

The stress-test note is on target. The parameter-free claim requires that the underlying MD force field already captures hotspot nucleation, growth, and the deflagration-to-detonation physics correctly, and that the trained network generalizes when the FE-scale microstructures differ in void statistics or binder fraction from the training set. Both conditions are weak points for classical potentials and for convolutional networks outside their training distribution. The abstract supplies no error metrics, no held-out validation, and no comparison to direct simulation or experiment, so the central claim cannot be assessed yet.

This work is aimed at modelers who need practical tools for energetic-materials safety and design. A reader already working on MD-to-continuum bridges would get concrete value from the training setup and coupling details once they are shown. The paper shows clear engagement with the multiscale problem and the literature, so it deserves a serious referee to examine the methods and results in full.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces MISTnetX, a convolutional deep neural network trained on molecular dynamics (MD) simulations of shock propagation through complex microstructures. The network supplies sub-grid source terms for shock-microstructure interactions, hotspot formation, and deflagration-to-detonation transition to continuum finite-element (FE) models. Applied to a synthetic nanostructured plastic-bonded RDX composite, the approach is claimed to enable parameter-free prediction of the full run-to-detonation transition by directly bridging atomistic and macroscale without traditional scale-separation assumptions.

Significance. If validated, the work would offer a machine-learning route to multiscale modeling of energetic materials where conventional methods fail due to absent scale separation. The direct use of MD-trained networks for sub-grid FE terms addresses a long-standing challenge in shock physics, though the parameter-free framing and generalization rest on unverified assumptions about MD fidelity and network extrapolation.

major comments (2)

[Abstract] Abstract: the central claim that MISTnetX enables 'parameter-free prediction' of the run-to-detonation transition is load-bearing yet unsupported, because the network is trained directly on MD outputs; this reduces the prediction to a surrogate mapping whose independence from the training data and force-field limitations is not demonstrated.
[Abstract] Abstract: no validation metrics, error analysis, training details, or tests of generalization to microstructures with different void statistics or binder fractions are supplied, leaving the weakest assumption (MD force-field accuracy for hotspot nucleation/growth and NN extrapolation beyond the training ensemble) unexamined and undermining the bridging claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments on the abstract. We respond to each point below and indicate where revisions will be made.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that MISTnetX enables 'parameter-free prediction' of the run-to-detonation transition is load-bearing yet unsupported, because the network is trained directly on MD outputs; this reduces the prediction to a surrogate mapping whose independence from the training data and force-field limitations is not demonstrated.

Authors: We agree the phrasing 'parameter-free' is imprecise and could suggest independence from the MD data and force field. The intent is to indicate that the continuum FE model contains no additional empirical parameters for sub-grid hotspot or DDT processes, as these are supplied by the MD-trained network rather than fitted at the macroscale. We will revise the abstract to clarify this distinction and note the dependence on the underlying MD force field. revision: yes
Referee: [Abstract] Abstract: no validation metrics, error analysis, training details, or tests of generalization to microstructures with different void statistics or binder fractions are supplied, leaving the weakest assumption (MD force-field accuracy for hotspot nucleation/growth and NN extrapolation beyond the training ensemble) unexamined and undermining the bridging claim.

Authors: The body of the manuscript reports validation metrics, training details, and error analysis (Sections 3–4) along with tests on held-out microstructures that vary in void size and binder fraction. We acknowledge these elements are not summarized in the abstract and that explicit discussion of force-field limitations and extrapolation would strengthen the bridging claim. We will add a concise statement of key metrics and generalization results to the abstract and expand the discussion of MD fidelity in the main text. revision: partial

Circularity Check

0 steps flagged

No significant circularity in the multiscale surrogate workflow

full rationale

The paper describes training a convolutional network (MISTnetX) on MD simulation outputs of shock propagation through microstructures, then using the trained network to supply sub-grid source terms inside FE simulations of a larger-scale composite. This is a standard data-driven surrogate approach for scale bridging; the final run-to-detonation prediction is the output of the FE solver driven by the network, not a quantity that reduces to the training data by definition or statistical forcing. No self-definitional equations, load-bearing self-citations, or renamings of known results appear in the derivation chain. The method remains self-contained against external benchmarks of MD fidelity and network generalization.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Abstract-only; full ledger cannot be populated. The central claim rests on the quality of MD training data and generalization of the neural network.

free parameters (1)

MISTnetX network parameters
Weights and biases fitted during training on MD data.

axioms (1)

domain assumption MD simulations of shock propagation through microstructures are sufficiently accurate and complete to serve as training data for the continuum scale.
Invoked implicitly as the source of all training information.

invented entities (1)

MISTnetX no independent evidence
purpose: Convolutional neural network surrogate for shock-microstructure interactions
New model introduced to bridge the scales.

pith-pipeline@v0.9.1-grok · 5685 in / 1159 out tokens · 51170 ms · 2026-06-29T16:33:29.180052+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 2 canonical work pages

[1]

Ignition and Growth Modeling of Shock Initiation of the TATB-based Explosives LX-17 and PBX 9502 at Eight Initial Temperatures Spanning a 446K Range

M. P. Kroonblawd, L. E. Fried, High Explosive Ignition through Chemically Activated Nanoscale Shear Bands. Phys. Rev. Lett. 124, 206002 (2020). 10. M. A. Wood, M. J. Cherukara, E. M. Kober, A. Strachan, Ultrafast Chemistry under Nonequilibrium Conditions and the Shock to Deflagration Transition at the Nanoscale. J. Phys. Chem. C 119, 22008–22015 (2015). 1...

work page arXiv 2020
[2]

Condensed Matter Detonation: Theory and Practice

P. C. H. Nguyen, Y.-T. Nguyen, J. B. Choi, P. K. Seshadri, H. S. Udaykumar, S. S. Baek, PARC: Physics-aware recurrent convolutional neural networks to assimilate meso scale reactive mechanics of energetic materials. Sci. Adv. 9 (2023). 23. P. C. H. Nguyen, X. Cheng, S. Azarfar, P. Seshadri, Y. T. Nguyen, M. Kim, S. Choi, H. S. Udaykumar, S. Baek, PARCv2: ...

work page doi:10.1007/978-3-642-22967-1_6 2023
[3]

Shock Initiation Experiments on PBX9501 Explosive at 150°C for Ignition and Growth Modeling

J. Xing, H. Zhang, L. Bai, G. Zhu, Q. Yu, B. Huang, Y. Liu, W. Wang, S. Li, Y. Liu, Nano-Voids in Ultrafine Explosive Particles: Characterization and Effects on Thermal Stability. Langmuir 39, 3391–3399 (2023). 36. J. D. Yeager, L. A. Kuettner, A. L. Duque, L. G. Hill, B. M. Patterson, Microcomputed X-Ray Tomographic Imaging and Image Processing for Micro...

2023
[4]

(44) 𝑭=𝐹4𝑭

M. L. Wilkins, Use of artificial viscosity in multidimensional fluid dynamic calculations. J. Comput. Phys. 36, 281–303 (1980). 47. O. Sen, P. K. Seshadri, N. K. Rai, J. Larentzos, J. Brennan, T. Sewell, C. R. Picu, H. S. Udaykumar, Johnson–Cook yield functions for cyclotetramethylene-tetranitramine (HMX) and cyclotrimethylene-trinitramine (RDX) derived f...

1980

[1] [1]

Ignition and Growth Modeling of Shock Initiation of the TATB-based Explosives LX-17 and PBX 9502 at Eight Initial Temperatures Spanning a 446K Range

M. P. Kroonblawd, L. E. Fried, High Explosive Ignition through Chemically Activated Nanoscale Shear Bands. Phys. Rev. Lett. 124, 206002 (2020). 10. M. A. Wood, M. J. Cherukara, E. M. Kober, A. Strachan, Ultrafast Chemistry under Nonequilibrium Conditions and the Shock to Deflagration Transition at the Nanoscale. J. Phys. Chem. C 119, 22008–22015 (2015). 1...

work page arXiv 2020

[2] [2]

Condensed Matter Detonation: Theory and Practice

P. C. H. Nguyen, Y.-T. Nguyen, J. B. Choi, P. K. Seshadri, H. S. Udaykumar, S. S. Baek, PARC: Physics-aware recurrent convolutional neural networks to assimilate meso scale reactive mechanics of energetic materials. Sci. Adv. 9 (2023). 23. P. C. H. Nguyen, X. Cheng, S. Azarfar, P. Seshadri, Y. T. Nguyen, M. Kim, S. Choi, H. S. Udaykumar, S. Baek, PARCv2: ...

work page doi:10.1007/978-3-642-22967-1_6 2023

[3] [3]

Shock Initiation Experiments on PBX9501 Explosive at 150°C for Ignition and Growth Modeling

J. Xing, H. Zhang, L. Bai, G. Zhu, Q. Yu, B. Huang, Y. Liu, W. Wang, S. Li, Y. Liu, Nano-Voids in Ultrafine Explosive Particles: Characterization and Effects on Thermal Stability. Langmuir 39, 3391–3399 (2023). 36. J. D. Yeager, L. A. Kuettner, A. L. Duque, L. G. Hill, B. M. Patterson, Microcomputed X-Ray Tomographic Imaging and Image Processing for Micro...

2023

[4] [4]

(44) 𝑭=𝐹4𝑭

M. L. Wilkins, Use of artificial viscosity in multidimensional fluid dynamic calculations. J. Comput. Phys. 36, 281–303 (1980). 47. O. Sen, P. K. Seshadri, N. K. Rai, J. Larentzos, J. Brennan, T. Sewell, C. R. Picu, H. S. Udaykumar, Johnson–Cook yield functions for cyclotetramethylene-tetranitramine (HMX) and cyclotrimethylene-trinitramine (RDX) derived f...

1980