Search for: All records

While machine learning (ML) interatomic potentials (IPs) are able to achieve accuracies nearing the level of noise inherent in the first-principles data to which they are trained, it remains to be shown if their increased complexities are strictly necessary for constructing high-quality IPs. In this work, we introduce a new MLIP framework which blends the simplicity of spline-based MEAM (s-MEAM) potentials with the flexibility of a neural network (NN) architecture. The proposed framework, which we call the spline-based neural network potential (s-NNP), is a simplified version of the traditional NNP that can be used to describe complex datasets in a computationally efficient manner. We demonstrate how this framework can be used to probe the boundary between classical and ML IPs, highlighting the benefits of key architectural changes. Furthermore, we show that using spline filters for encoding atomic environments results in a readily interpreted embedding layer which can be coupled with modifications to the NN to incorporate expected physical behaviors and improve overall interpretability. Finally, we test the flexibility of the spline filters, observing that they can be shared across multiple chemical systems in order to provide a convenient reference point from which to begin performing cross-system analyses. 

The melting temperature is important for materials design because of its relationship with thermal stability, synthesis, and processing conditions. Current empirical and computational melting point estimation techniques are limited in scope, computational feasibility, or interpretability. We report the development of a machine learning methodology for predicting melting temperatures of binary ionic solid materials. We evaluated different machine-learning models trained on a dataset of the melting points of 476 non-metallic crystalline binary compounds using materials embeddings constructed from elemental properties and density-functional theory calculations as model inputs. A direct supervised-learning approach yields a mean absolute error of around 180 K but suffers from low interpretability. We find that the fidelity of predictions can further be improved by introducing an additional unsupervised-learning step that first classifies the materials before the melting-point regression. Not only does this two-step model exhibit improved accuracy, but the approach also provides a level of interpretability with insights into feature importance and different types of melting that depend on the specific atomic bonding inside a material. Motivated by this finding, we used a symbolic learning approach to find interpretable physical models for the melting temperature, which recovered the best-performing features from both prior models and provided additional interpretability. 

Interstitial dumbbell-mediated diffusion can affect segregation and precipitation properties of solutes in alloys under irradiated conditions. Accurate computation of transport coefficients for dumbbell-mediated diffusion thus becomes essential for modelling solute segregation under irradiation. In this work, we extend the Green’s function approach, a general numerical approach, to compute accurate transport coefficients for interstitial dumbbell-mediated mechanisms in the dilute limit for arbitrary crystalline systems with non-truncated correlations in atomic diffusion. We also present results of tracer correlation factors, solute drag ratios and partial diffusion coefficient ratios in iron and nickel-based alloys computed with our approach, compare our results with existing results in the literature, and discuss some aspects of correlated solute-dumbbell motion. 

Computational methods have gained importance and popularity in both academia and industry for materials research and development in recent years. Since 2014, our team at University of Illinois at Urbana-Champaign has consistently worked on reforming our Materials Science and Engineering curriculum by incorporating computational modules into all mandatory undergraduate courses. The outbreak of the COVID-19 pandemic disrupted education as on-campus resources and activities became highly restricted. Here we seek to investigate the impact of the university moving online in Spring 2020 and resuming in-person instructions in Fall 2021 on the effectiveness of our computational curricular reform from the students' perspective. We track and compare feedback from students in a representative course MSE 182 for their computational learning experience before, during and after the pandemic lockdown from 2019 to 2021. Besides, we survey all undergraduate students, for their online learning experiences during the pandemic. We find that online learning enhances the students' belief in the importance and benefits of computation in materials science and engineering, while making them less comfortable and confident to acquire skills that are relatively difficult. In addition, early computational learners are likely to experience more difficulties with online learning compared to students at late stages of their undergraduate education, regardless of the computational workload. Multiple reasons are found to limit the students' online computational learning, such as insufficient support from instructors and TAs, limited chances of peer communication and harder access to computational resources. Therefore, it is advised to guarantee more resources to students with novice computational skills regarding such limiting reasons in the future when online learning is applied. 

We present an overview of four challenging research areas in multiscale physics and engineering as well as four data science topics that may be developed for addressing these challenges. We focus on multiscale spatiotemporal problems in light of the importance of understanding the accompanying scientific processes and engineering ideas, where “multiscale” refers to concurrent, non-trivial and coupled models over scales separated by orders of magnitude in either space, time, energy, momenta, or any other relevant parameter. Specifically, we consider problems where the data may be obtained at various resolutions; analyzing such data and constructing coupled models led to open research questions in various applications of data science. Numeric studies are reported for one of the data science techniques discussed here for illustration, namely, on approximate Bayesian computations. 

Over the past years, our team has taken a concerted effort to integrate computational modules into courses across the undergraduate curriculum, in order to equip students with computational skills in a variety of contexts that span the field of Materials Science and Engineering. This effort has proven sustainable during the recent period of online transition of many courses, illustrating one of the benefits of computational modules. The most recent addition to our set of modules included a visualization component that was incorporated into our introductory freshman course for the first time in Fall 2019. Students can perform this module either using local computing labs, access those resources remotely, or can use their own computers. In the Fall of 2020, we modified this module and expanded it towards the utilization of a materials database to teach students how to search for materials with specific properties. The results were then interfaced with the previously existing visualization module to connect the structure and symmetry of materials with their properties and to compare them with experimental results. We implement a more detailed survey to learn to what extent students gained the capability of using databases for future research and education. We will also use these responses to further develop and improve our existing modules.