Session K02: Machine Learning for Energetic Materials I

Mapping microstructure to shock-induced temperature fields using machine learning

The response of materials to dynamical, or shock, loading is important to planetary science, aerospace engineering, and energetic materials. Thermal-activated processes, including chemical reactions and phase transitions, are significantly accelerated by the localization of the energy deposited into hotspots. These result from the interaction of a supersonic wave with the materials’ microstructure and are governed by the collapse of porosity, interfacial friction, and localized plastic deformation. These mechanisms are not fully understood and today we lack predictive models to, for example, predict the shock to detonation transition chemistry and microstructure alone. We demonstrate that deep learning techniques can be trained to predict the resulting temperature fields from large-scale molecular dynamics simulations with the initial microstructure as the only input in complex polymer composite systems. The model accuracy is enough for quantitative prediction of the initiation of chemical reactions.   

Predicting critical impact velocity in PBX-9501 using machine learning

Heterogeneous energetic materials (HEMs) can involve microstructural defects such as randomly distributed pores of varying size and shape. These heterogeneities introduce temperature spikes known as hotspots which affect their shock response. Current methods for obtaining detonation properties rely on experiments and computational models. However, accounting for each possible pore configuration requires an extensive number of experiments and computational simulations, making them unfeasible for this problem. Machine Learning (ML) offers an attractive approach to overcome these challenges. Towards this goal, we develop a ML model for predicting critical velocities in PBX-9501 samples with multiple pores of varying quantity, size, and spatial distribution. We used CTH to emulate the shock response of each sample when impacted by a flyer plate. To compute the resulting critical velocities, an automated bisection algorithm was developed. We then leveraged three ML models: Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Graph Neural Network (GNN). The performance of each model, as they compare in prediction of critical velocity, is studied. The models are stress-tested for cases involving new spatial distributions, pore quantities, and pore sizes. It is expected that these models will directly predict the critical velocity for unseen pore structures without executing CTH. Ultimately, we believe this work paves the way towards ML-guided models to understand how various pore structures affect shock sensitivity in HEMs   

“Research output software for energetic materials based on observational modelling/ machine learning” (RoseBoom©)

There is huge scope for the implementation of sustainable methods in the research of new energetic materials. It is certainly one of the most important aspects which must be considered and implemented in current and future modern scientific research. There are a number of ways this can be achieved, and with the development of the program “Research output software for energetic materials based on observational modelling/ machine learning” (RoseBoom^©) it is hoped that the development of new modern energetic materials will be advanced, since it aims to provide access to quick and easy prediction methods which will indicate performance parameters (e.g. the detonation velocity and pressure, the key indicator for the power of an explosive) – before they have been synthesized. The software allows fast estimation of the performance, enthalpy of formation and density of new energetic compounds only based on the structural formula. To do this it combines empirical and machine learning models into one program, that can be used to evaluate performance of new energetic materials before synthesis and after synthesis within experimental uncertainty. The user-friendly design allows fast computation of hundreds of molecules within a few minutes with minimal user-input. A picture of a compound is sufficient, which can be taken using the screenshot function implemented in RoseBoom^©, the molecule can be copied from a molecule editor, or a list of molecules/mixtures can be loaded into the program, obtaining the results in an Excel spreadsheet.   

Group Equivalent Machine Learning Approach to Predict Hydrocarbon Strain Energy

Strain energy is an important property in molecules intended for reactive and energetic applications. We present a machine learning approach to predict hydrocarbon strain energies using Benson group equivalents. An algorithm is developed to break down hydrocarbons into their group equivalent components and a combined featurization strategy is developed using the group equivalents and other simple physicochemical features. The training data are obtained from a limited number of quantum chemistry simulations. A machine learning approach to predict hydrocarbon molecule strain energies is then described and evaluated.

    

Automatic Differentiation in Dynamic Topology Optimization

Recent advances in scientific programming, particularly with regards to topology optimization and machine learning, have necessitated computational methods that generate gradients for optimization. One method to do so is to utilize a tool called automatic differentiation, a mechanism to algorithmically calculate derivatives of functions and combine them to generate gradients of compositions of functions. Code bases such as Jax and PyTorch (which particularly focused on machine learning applications) have demonstrated the ability to scale automatic differentiation to large problems. This allows for rapid gradient calculations, leading to reduced development time as well as significantly higher complexity in the equations used to study a physical phenomenon. While these methods have been studied in the context of machine learning, these approaches have only been applied to mechanics in a handful of cases. This presents an opportunity to study a large variety of optimization problems, such as topology, material parameters, or initial-value problems using this automatic differentiation infrastructure. By taking advantage of the sequential structure of time dependent problems, we develop rapid algorithms for gradient calculations that can be utilized in a large variety of different contexts. We demonstrate the efficacy of these algorithms by studying various dynamics problems. One example is in designing the initial velocity profile of an elastic wave to generate a specific final state. This can be utilized to amplify or reduce particular wave characteristics in non-trivial ways. This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.