Building Chemical Property Models for Energetic Materials from Small Datasets Using a Transfer Learning Approach.

For many experimentally measured chemical properties that cannot be directly computed from first-principles, the existing physics-based models do not extrapolate well to out-of-sample molecules, and experimental datasets themselves are too small for traditional machine learning (ML) approaches. To overcome these limitations, we apply a transfer learning approach, whereby we simultaneously train a multi-target regression model on a small number of molecules with experimentally measured values and a large number of molecules with related computed properties. We demonstrate this methodology on predicting the experimentally measured impact sensitivity of energetic crystals, finding that both characteristics of the computed dataset and model architecture are important to prediction accuracy of the small experimental dataset. Our directed-message passing neural network (D-MPNN) ML model using transfer learning outperforms direct-ML and physics-based models on a diverse test set, and the new methods described here are widely applicable to modeling many other structure-property relationships.

Spectroscopic Probe Molecule Selection Using Quantum Theory, First-Principles Calculations, and Machine Learning.

Probe molecule vibrational spectra have a long history of being used to characterize materials including metals, oxides, metal-organic frameworks, and even human proteins. Furthermore, recent advances in machine learning have enabled computationally generated spectra to aid in detailed characterization of complex surfaces with probe molecules. Despite widespread use of probe molecules, the science of probe molecule selection is underdeveloped. Here, we develop physical concepts, including orbital interaction energy and the energy overlap integral, to explain and predict the ability of probe molecules to discriminate structural descriptors. We resolve the crystal orbital overlap population (COOP) to specific molecular orbitals and quantify their bonding character, which directly influences vibrational frequencies. Using only a single adsorbate calculation from density function theory (DFT), we compute the interaction energy of individual adsorbate molecular orbitals with adsorption site atomic orbitals across many different sites. Combining the molecular orbital resolved COOP and changes in orbital interaction energy enables probe molecule selection for improved discrimination of various sites. We demonstrate these concepts by comparing the predicted effectiveness of carbon monoxide (CO), nitric oxide (NO), and ethylene (C2H4) to probe Pt adsorption sites. Finally, using a previously developed machine learning framework, we show that models trained on hundreds of thousands of C2H4 spectra, computed from DFT, which regress surface binding-type and generalized coordination number, outperform those trained using CO and NO spectra. A python package, pDOS_overlap, for implementing the electron density-based analysis on any combination of adsorbates and materials, is also made available.

Explainable and trustworthy artificial intelligence for correctable modeling in chemical sciences.

Data science has primarily focused on big data, but for many physics, chemistry, and engineering applications, data are often small, correlated and, thus, low dimensional, and sourced from both computations and experiments with various levels of noise. Typical statistics and machine learning methods do not work for these cases. Expert knowledge is essential, but a systematic framework for incorporating it into physics-based models under uncertainty is lacking. Here, we develop a mathematical and computational framework for probabilistic artificial intelligence (AI)-based predictive modeling combining data, expert knowledge, multiscale models, and information theory through uncertainty quantification and probabilistic graphical models (PGMs). We apply PGMs to chemistry specifically and develop predictive guarantees for PGMs generally. Our proposed framework, combining AI and uncertainty quantification, provides explainable results leading to correctable and, eventually, trustworthy models. The proposed framework is demonstrated on a microkinetic model of the oxygen reduction reaction.

Infrared spectroscopy data- and physics-driven machine learning for characterizing surface microstructure of complex materials.

There is a need to characterize complex materials and their dynamics under reaction conditions to accelerate materials design. Adsorbate vibrational excitations are selective to adsorbate/surface interactions and infrared (IR) spectra associated with activating adsorbate vibrational modes are accurate, capture details of most modes, and can be obtained operando. Current interpretation depends on heuristic peak assignments for simple spectra, precluding the possibility of obtaining detailed structural information. Here, we combine data-based approaches with chemistry-dependent problem formulation to develop physics-driven surrogate models that generate synthetic IR spectra from first-principles calculations. Using synthetic IR spectra of carbon monoxide on platinum, we implement multinomial regression via neural network ensembles to learn probability distributions functions (pdfs) that describe adsorption sites and quantify uncertainty. We use these pdfs to infer detailed surface microstructure from experimental spectra and extend this methodology to other systems as a first step towards characterizing complex interfaces and closing the materials gap.

Scaling relationships and theory for vibrational frequencies of adsorbates on transition metal surfaces.

Adsorbate vibrational excitations are an important fingerprint of molecule/surface interactions, affecting temperature contributions to the free energy and impacting reaction rate and equilibrium constants. Furthermore, vibrational spectra aid in identifying species and adsorption sites present in experimental studies. Despite their importance, knowledge of how adsorbate frequencies scale across materials is lacking. Here, by combining previously reported experimental data and our own density-functional theory calculations, we reveal linear correlations between vibrational frequencies of adsorbates on transition metal surfaces. Through effective-medium theory, linear muffin-tin orbital theory, and the d-band model, we rationalize the squares of the frequencies to be fundamentally linear in their scaling across transition metal surfaces. We identify the adsorbate-binding energy as a descriptor for certain molecular vibrations and rigorously relate errors in frequencies to errors in adsorption energies. We also discuss the impact of scaling on surface thermochemistry and adsorbate coverage.