Locally Optimizable Joint Embedding Framework to Design Nitrogen‐rich Molecules that are Similar but Improved

Abstract

Deep learning has shown great potential for generating molecules with desired properties. But the cost and time required to obtain relevant property data have limited study to only a few classes of materials for which extensive data have already been collected. We develop a deep learning method that combines a generative model with a property prediction model to fuse small data of one class of molecules with larger data in another class. Common low‐level physicochemical properties are jointly embedded into a latent space that can be used to design molecules in the smaller class. The chemical space around the molecules in the training set is explored through local gradient ascent optimization. Based on nine molecules from the original training set, nine new molecules are found to have improved properties while remaining structurally similar to the training molecules thereby easing requirements for entirely new synthesis routes. Validation is performed using an equilibrium thermochemistry code to verify the molecules and target properties. A specific example targeting the Chapman‐Jouguet velocity and small data for nitrogen‐rich molecules is shown. Despite the relative lack of nitrogen‐rich molecule data, the results demonstrate that fusing and joint embedding with plentiful low nitrogen molecular data can produce higher generative performance than using the scarce data alone.

Releasing chemical energy in spatially programmed ferroelectrics

“…MAE is averaged over the three random train test splits, and the best model and featurization scheme that gives the minimum MAE is chosen. Autoencoder in the joint embedding framework is trained with 100 K molecules from MOSES 40 in addition to the entire set of 401 energetic molecules 35 . Finally, the best model and featurization scheme chosen from model cross validation process is used to predict the detonation velocity of each ferroelectric material in the dataset.…”

“…The eight models include Gaussian Process Regression, Kernel Ridge Regression, Support Vector Regression, Random Forest, Lasso Regression, k-Nearest Neighbors, Gradient Boosting and Ridge Regression. Previous studies on structure property prediction in energetic materials have shown good performance with these models 35,36 . Hyperparameter optimization is performed on each model (and for each featurization) using grid search with 5-fold cross validation.…”

“…Five different featurization methods are used in this paper: Sum Over Bonds (SoB), E State, a Custom Descriptor Set (CDS), Joint Embedding 35 , and a concatenation of Sum Over Bonds, E State and Customer Descriptor Set 22 . The set of features was shown earlier to result in less mean absolute error (MAE) and improved performance for the estimation of exothermic reaction parameters 35,36 . Eight separately trained machine learning models are examined as property predictors, that takes an input feature and predict detonation velocity as the output.…”

Artificial Intelligence Approaches for Energetic Materials by Design: State of the Art, Challenges, and Future Directions

“…Once they have been rigorously validated, computational forward design models allow us to not only quantify properties and performance but also to understand the underlying mechanics of the ignition and combustion process, such as the crucial hotspot physics and its connection to the microstructure [1,16]. In recent years sev-eral research groups have furthered understanding of the relationships among structure, property, and performance of a range of CHNO EMs [44][45][46][47]. These studies have revealed important relationships among porosity [1], pore size [1,48], shape and distribution [15,45,[49][50][51], and other morphological features such as asperities and interfaces between crystals and binders [46,52], and their effects on the sensitivity of the composite material.…”

Predicting Enthalpy of Formation of Energetic Compounds by Machine Learning: Comparison of Featurization Methods and Algorithms

“…In the field of energetic materials, the availability of a complete public dataset is still the bottleneck for building ML models. In most studies, researchers chose to gather data from literature [12][13][14] and extracted them from public databases [15][16][17], while a few others used results from theoretical calculations [18,19]. In this study, we extracted data from the literature.…”