Machine Learning for Discovery of Structure-Process-Property Relations in Electronic Materials: Machine Learning in Materials Engineering I
Sponsored by: ACerS Electronics Division
Program Organizers: B. Reeja Jayan, Carnegie Mellon University; Aarti Singh, Carnegie Mellon University
Monday 8:00 AM
November 2, 2020
Room: Virtual Meeting Room 43
Location: MS&T Virtual
Session Chair: B. Reeja Jayan, Carnegie Mellon University
8:00 AM
Introductory Comments: Machine Learning for Discovery of Structure-Process-Property Relations in Electronic Materials: B. Reeja Jayan1; 1Carnegie Mellon University
Introductory Comments
8:05 AM
Expert-guided Learning for Data-constrained Materials Science Problems: Gopaljee Atulya1; Shuyan Zhang1; Alexander Davis1; 1Carnegie Mellon University
Machine learning approaches are most successful when data is abundant to train models. In these cases, algorithms learn complicated patterns without requiring significant domain knowledge from experts. However, many materials science problems are complex (high dimensional) and data are expensive to collect, but experts have extensive knowledge that can augment purely algorithmic approaches. We discuss a framework for incorporating expert feedback in learning and optimizing algorithms for materials science problems. We combine concepts from choice modeling with machine learning models to produce novel algorithms that utilize an expert’s domain expertise as a source of data in conjunction with simulations and experiments. We explore the types of queries that extract the most information from experts when trying to learn a mapping while minimizing the cognitive burden on experts. We apply this framework to solve problems in identifying and analyzing pair distribution functions (PDF) of experimentally synthesized ceramics specifically titanium dioxide (TiO2).
8:25 AM
Parametric Analysis to Quantify Process Input Influence on the Printed Densities of Binder Jetted Alumina Ceramics: Edgar Mendoza Jimenez1; Jack Beuth1; Baby Reeja-Jayan1; 1Carnegie Mellon University
Binder jetting can become a viable method to additively manufacture ceramics. However, the effects of process parameters/inputs on printing outputs (e.g. part density and geometric resolution) have not been investigated for the binder jetting of ceramic powder systems. In this work, a parametric study explores the influence of seven process inputs on the relative densities of as-printed (green) alumina parts. Sensitivity analyses compare the influence of each input on green densities. Multivariable linear and Gaussian process regression provide models for predicting green densities as a function of binder jetting process inputs. The models indicate that the green densities of alumina builds can be increased by decreasing the recoat speed and increasing the oscillator speed and the rest of the parameters have a nonlinear influence. Results reported in this study can be leveraged to control the porosity of binder jetted parts for applications such as filters, bearings, electronics, and medical implants.
8:45 AM
SimuLearn: Machine Learning-empowered Fast and Accurate Simulator to Support 4D Printing Design: Humphrey Yang1; Kuanren Qian1; Haolin Liu1; Yuxuan Yu1; Jianzhe Gu1; Matthew McGehee1; Yongjie Zhang1; Lining Yao1; 1Carnegie Mellon University
Recent technological advancements have created a library of smart materials that afford novel functionalities. 4D printing, in particular, administers more efficient and economical prototyping as well as manufacturing. However, due to the lack of fast and accurate transformation simulators, currently available 4D printing CAD tools cannot effectively support users to iterate designs that have complex topologies. To address this issue, we take mesh-like structures as an example to introduce a novel SimuLearn system that combines finite element analysis (FEA) and graph convolutional networks (GCN) to truthfully (97% accuracy versus FEA) inform design decisions in real-time (0.6 seconds) and deploy our implementation in a computational design tool to unveil the enabled design space. Results show that SimuLearn enables much faster design iteration and allows users to integrate material response into their design workflows. Beyond 4D printing, SimuLearn also enriches the computational toolbox for designing, engineering, and predicting smart, transformative materials.
9:05 AM
Neural Network Potential for Lattice Dynamics Calculations and Thermal Conductivity Prediction: Jie Gong1; Hyun-Young Kim1; Alan McGaughey1; 1Carnegie Mellon University
Lattice dynamics calculations can be used to predict the phonon properties of insulating and semi-conducting crystals. These calculations require force constants, which can be found using density functional theory (DFT). The force constants of simple materials (high symmetry and small primitive cell) can be found with relatively few DFT calculations, but this number increases significantly for more complex materials, taxing computational resources.We address this issue by training a high-dimensional neural network potential to calculate force constants with a small training set. An adaptive selection scheme is used to select the training data efficiently. Using silicon, we quantify the accuracy using the phonon frequencies and thermal conductivity, in addition to the standard force and energy metrics. We find that accurate forces, energies, and frequencies do not guarantee an accurate thermal conductivity. A single training set and hyperparameters can result in a range of thermal conductivities.
9:25 AM
Uncertainty Quantification and Active Learning of Neural Network Models for Predicting ZrO2 Crystal Energy: Jayanth Koushik1; Sungjun Choi1; Aarti Singh1; 1Carnegie Mellon University
Neural networks are increasingly being used to model complex functions that arise naturally in material science; networks trained to predict crystal energies can avoid prohibitively expensive computations. However, it is challenging to analyze predictions of neural networks to guide further analysis because the prediction mechanism is poorly understood. One issue is obtaining uncertainty estimates of predictions, which can be used to identify abnormal data points, or adaptively sample additional points. We present a novel algorithm to efficiently approximate predictive variances of neural networks. Our method uses the same idea as the Jackknife method from statistics, but avoids any re-training, making it scalable to large datasets. We apply our method on a network trained to predict energy of ZrO2 crystals, and successfully identify mislabeled and abnormal structures in the data set. We also demonstrate improved performance in training the network actively, when points are sampled based on uncertainty rather than randomly.