ICME Gap Analysis in Materials Informatics: Databases, Machine Learning, and Data-Driven Design: Session IV
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: James Saal, Citrine Informatics; Carelyn Campbell, National Institute of Standards and Technology; Raymundo Arroyave, Texas A&M University

Thursday 2:00 PM
February 27, 2020
Room: 30D
Location: San Diego Convention Ctr

Session Chair: James Saal, Citrine Informatics


2:00 PM  
Improved Performance of Automatic Characterization of Steel Microstructure by Machine Learning Architecture: Jonghyuk Lee1; Seonghwan Kim1; Nam Hoon Goo1; 1Hyundai Steel
    The microstructure of steel is a significant factor in determining the performance of the material. Despite this importance, it is still challenging to quantitatively and systematically interpret the microstructure due to the multiple phase changes and the various precipitates. The manual classification of the microstructural images, which vary widely depending on the alloying elements and the process conditions, has limitation in reproducibility and accuracy. Recent studies show that the machine learning algorithm like FCN (Fully Convolutional Network) with the featured segmentation has excellent performance over 90% match. We have implemented a machine learning approach in the analysis of steel microstructures. We have analyzed the microstructure of steel using FCN and Pix neural networks, which are recently known to show excellent performance, and compared the results of two neural networks. The approach eliminates uncertainty due to human error, and the analysis results reflect realistic mechanical properties of actual steel material.

2:20 PM  
Training Data-driven Machine Learning Models Using Physics Simulations: Predicting Local Thermal Histories in Additive Manufactured Components: Michael Groeber1; Karthik Giriprasad1; 1The Ohio State University
     The processing parameter space in additive manufacturing (AM) is prohibitively large – driving the need for ICME processing tools. In powder bed fusion processes (e.g. LPBF or EBM) printing parameters such as beam power, focus, raster speed, and raster path affect thermal and mechanical stress states. Mapping influences of these parameters is daunting, preventing design of customized microstructures theoretically achievable via AM.Energy input time series, listing relative distances, timings and intensities of energy input, represent the local scan path at locations in AM components. We investigate processing these time series through a 1D, deep convolutional neural network (CNN) to predict local thermal histories. A fast-acting analytical model predicts thermal histories for training the CNN. We aim to show even efficient thermal models can be replaced by machine learning models, derived from novel representations of local energy input. These models provide additional promise when trained against advanced physics.

2:40 PM  
Relating Microstructure Features to Response Using Convolutional Neural Networks: Sean Donegan1; Navneet Kumar2; Michael Groeber2; 1Air Force Research Laboratory; 2The Ohio State University
    Fast acting methods for quantifying the relationship between microstructure and properties are an enabling capability for ICME-driven materials design. Given a property of interest, a number of microstructural factors may contribute to the magnitude of its response. For example, hot spots in the local stress state of a polycrystal may arise from elastic anisotropy, local morphology, and crystallographic orientations relative to the load state. It is therefore advantageous to investigate frameworks that predict properties directly from an input microstructure, learning the necessary features based on training data. We describe such a method using convolutional neural networks (CNNs), parameterizing the input microstructure as an image. The CNN is trained by creating synthetic microstructures in DREAM.3D whose elastic response is modeled using a technique based on fast Fourier transforms. We discuss the importance of capturing key information, such as crystallographic anisotropy, in the microstructural image description, and its impact on model accuracy.

3:00 PM  
Prediction of Steel Micro-structure by Deep Learning Using Database of Thermo-dynamics and Phase Field Model: Seonghwan Kim1; Hyeok Jae Jeong1; Jong Hyuk Lee1; Nam Hoon Goo1; 1Hyundai Steel Company
    Data-driven design of materials is getting attention as method for searching materials with improvement of machine learning. Therefore, a database for design is very important. As database, the real properties is best choice, but stacking real data as database is very hard work consuming large amount of time and cost. In addition, real data has high measurement error. Thus, ICME can be one of important alternative database. However, there are drawbacks in ICME scheme such long calculation times, defining various fitting parameters, and combining various tools. Thus, in order to overcome these disadvantages of ICME scheme, we suggest new method for prediction of micro-structure of steel, which is one of consuming part of ICME. Using two different type of deep neural networks of DNN regression and GAN trained by database of thermo-dynamics and phase field model, the prediction of micro-structure in accordance with processing parameters is achieved in several seconds.

3:20 PM  
Reduction of Uncertainty in a First-principles-based CALPHAD-type Phase Diagram via Sequential Learning of Phase Equilibrium Data: Theresa Davey1; Brandon Bocklund2; Zi-Kui Liu2; Ying Chen1; 1Tohoku University; 2Pennsylvania State University
     Phase diagrams are a fundamental tool in materials design, but thorough experimental determination is challenging, expensive, and time consuming. Phase diagrams calculated entirely from first-principles may reduce time and expense, providing information at the prediction stage. Our previous work demonstrated a methodology to obtain a first-principles-only CALPHAD-type phase diagram reproducing all major features, with little or no prior knowledge of the system [1]. This can guide reduced experiments needed for database validation. Considering the quantified uncertainty of the phase diagram [2] using ESPEI [3], a sequential learning approach is taken to systematically add data in regions of highest uncertainty. This models how the first-principles only phase diagram could help select experimental parameters, and how each experiment affects the phase diagram. [1] T. Davey et al., CALPHAD XLVIII, June 2019. [2] N. Paulson et al., Acta Mater. 174 (2019) 9–15.[3] B. Bocklund et al., MRS Commun. (2019) 1-10.

3:40 PM Break

4:00 PM  Cancelled
Artificial Materials Intelligence (AMI) to Accelerate Discovery of Novel Superalloys: Irina Roslyakova1; Setareh Zomorodpoosh1; Mansur Ahmed1; Abdulmonem Obaied1; Ingo Steinbach1; 1ICAMS, Ruhr-University Bochum
    A novel modeling strategy, which combines artificial intelligence (AI) with artificial materials (AM) will be proposed and called Artificial Materials Intelligence (AMI). Established physical laws, cross-correlations between different materials properties and well developed thermodynamic databases will be considered during the development and the application of AMI. Machine learning will be used if the established physical basis is insufficient for predictive materials models. The application of such combined methodology allows to keep as much physics as possible on the one side and reduce the number of exploratory variables and required data on the other side. The proposed study will demonstrate the applicability of high throughput simulations and physically-based data-driven modeling strategy combined in AMI to predict mechanical properties of complex, microstructure dominated materials, such as Ni- and Co-based superalloys and to accelerate discovery of optimal chemical composition for these alloys.

4:20 PM  
Steel Development and Optimization Using Response Surface Models: Jun Hu1; Rachael Stewart1; Erik Pavlina1; Grant Thomas1; Alexander Duggan2; Roel Van De Velde2; 1AK Steel; 2ESTECO
    With more advanced and sophisticated technologies implemented into steel development, conventional empirical methodology is becoming more inefficient and undirected in this process. Response surface models are thereby introduced into this field to integrate ‘big data’ and computationally bridge inputs to outputs. In this work, a completed procedure will be presented to show training response surface models using different algorithms based on a steel chemistry and processing database with corresponding mechanical properties. Furthermore, optimization will be applied to mine feasible but undeveloped new steel possibilities from the well-trained response surface model. To validate the computation, a laboratory steel is processed, and the resulting mechanical properties are and compared with the computational results.

4:40 PM  Cancelled
Utilizing the Statistical Machine Learning Approaches to Design New NiTiHf High Temperature Shape Memory Alloys: Tejas Umale1; Shahin Boluki1; Xiaoning Qian1; Raymundo Arroyave1; Ibrahim Karaman1; 1Texas A&M University
    Ni-Ti-Hf High Temperature Shape Memory Alloys (HTSMAs) are potential candidates for aerospace and space applications, due to their high strength, tailorable transformation temperatures, and high actuation energies. Change in initial composition and thermal processing can be utilized to achieve a wide range of TTs. In this work, we aim to utilize the compositional dependence of TTs data from our previous work, for designing novel NiTiHf alloys with a targeted martensitic start temperature (Ms) between 200°C to 300°C and minimum thermal hysteresis (Ap-Mp). A Bayesian Global Optimization Framework was implemented on the initial dataset to design new alloys with minimum number of iterative experiments. In the framework, Bayesian probabilistic models were created and various constraint optimization based selectors were implemented in order to make the predictions and new alloys were fabricated. Such an iterative framework, led us to discover minimum reported thermal hysteresis in literature while satisfying the Ms criteria.

5:00 PM  
Machine Learning-directed Navigation of Synthetic Design Space: A Statistical Learning Approach to Controlling the Synthesis of Perovskite Halide Nanoplatelets in the Quantum-confined Regime: Erick Braham1; Junsang Cho1; Kristel Forlano1; Raymundo Arroyave1; Sarbajit Banerjee1; 1Texas A&M University
    Utilizing the synthesis of two-dimensional CsPbBr3 nanoplatelets as a model system, we demonstrate an efficient machine learning navigation of reaction space that allows for predictive control of layer thickness down to sub-monolayer dimensions. Support vector machine (SVM) classification and regression models are used to initially separate regions of the design space that yield quantum-confined nanoplatelets from regions yielding bulk particles and subsequently to predict the thickness of quantum-confined CsPbBr3 nanoplatelets that can be accessed under specific reaction conditions. The SVM models are not only just predictive and efficient in sampling the available design space but also provide fundamental insight into the influence of molecular ligands in constraining the dimensions of nanocrystals. The results illustrate a quantitative approach for efficient navigation of reaction design space and pave the way to navigation of more elaborate landscapes beyond dimensional control spanning polymorphs, compositional variants, and surface chemistry.