Session Sheet - 3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Optimization of Manufacturing Processes II

3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Optimization of Manufacturing Processes II
Program Organizers: Remi Dingreville, Sandia National Laboratories; Ali Riza Durmaz, Fraunhofer Institute Iwm

Wednesday 1:40 PM
June 18, 2025
Room: Elite Ballroom 1 & 2
Location: Anaheim Marriott

Session Chair: Sung Park

1:40 PM
AI Guided Discovery of Lunar Derived Materials for a Sustainable Ecosystem: Eddie Gienger¹; Michael Pekala¹; Nam Le¹; Alex New¹; Greg Canal¹; Karun Kumar Rao¹; Milena Graziano¹; Morgan Trexler¹; Christian Sanjurjo-Rodriguez¹; Steven Storck¹; Elizabeth Pogue¹; Mary Daffron¹; Greg Bassen²; Aaron Baumgarten¹; Brandon Wilfong²; Denise Yin¹; Wyatt Bunstine²; Elizabeth Reilly¹; Leslie Hamilton¹; Tyrel McQueen²; Christopher Stiles¹; Christopher Stiles¹; ¹Johns Hopkins University - Applied Physics Laboratory; ²Johns Hopkins University
    One key challenge for lunar missions is manufacturing materials directly on the Moon using in situ resources. By integrating AI, high-throughput synthesis, and large language models (LLMs), we advance materials discovery, design, and fabrication in the Moon’s harsh environment, reducing dependence on Earth-based resources. The core “predict-make-measure” framework utilizes AI-driven predictions, experimental synthesis, and rapid characterization, creating a closed-loop system that iterates efficiently from material concept to production. An AI-powered knowledgebase enables efficient search and optimization of lunar resources. This database aids in discovering high-hardness, high-silicon materials composed of lunar-abundant elements, while generative models and a physics-based AI screening pipeline refine composition candidates. Utilizing directed energy deposition (DED) over 250 unique materials have been synthesized. The samples are characterized with quick assessments of mechanical properties like hardness and qualitative ductility scoring. This approach sets a foundation for autonomous, sustainable material synthesis in resource constrained environments.

2:00 PM
A Data-Driven Approach to Print Performance Prediction : Jennifer Ruddock¹; James Hardin¹; ¹Air Force Research Laboratory
    DIW 3D printing is a useful additive manufacturing technique for high-mix, low-volume manufacturing because the process parameters can be adapted to new materials and geometries. However, the link between printer parameters, ink properties and printed part performance is often not straightforward, resulting in a costly trial-and-error approach to print process adaptation. To avoid this onerous process, we aim to determine what minimal test print pattern can be used to predict useful print performance metrics such as the geometric fidelity (avoiding slumping, voids, etc.) and minimize overall print time. This work will first build a core dataset of potential test prints and representative print challenges then seek to link test prints to print behavior using combinations of data-driven tools.

2:20 PM
Simulation-Based Optimization of Additive Manufacturing Toolpaths to Reduce Distortion: Ashley Gannon¹; Stephen DeWitt¹; James Haley¹; Bruno Turcksin¹; Lauren Heinrich¹; Thomas Feldhausen¹; Alex Beatty¹; Alex Roschli¹; Andres Marquez Rossy¹; Leah Jacobs¹; Cameron Adkins¹; Callan Herberger; Callan Herberger¹; Michael Borish¹; Liam White¹; ¹Oak Ridge National Lab
    Minimizing residual stresses and the resulting geometric distortion is important for enabling a wider adoption of additively manufactured parts. The toolpath during printing, including the dwell time between layers, is known to have a significant impact on the residual stress. We present a workflow that combines toolpath planning, simulation, in-situ monitoring, and optimization techniques to refine dwell control for improved geometric accuracy. This workflow integrates finite element analysis with real-time monitoring data to iteratively calibrate and validate simulation predictions, forming a feedback system that minimizes distortion. To quantify deviations, we use a structured blue light scanner to capture the geometry of printed part and compare the scanned geometry to the original CAD. This comparison allows us to assess distortion quantitatively and verify the effectiveness of optimized dwell parameters in reducing distortion. Additionally, this workflow can be adapted for convergent manufacturing processes, enhancing the accuracy and efficiency of hybrid production methods.

2:40 PM
Utilizing Time-Series Data for Improved Prediction of End-Point Temperature and Carbon in Basic Oxygen Furnace With a Large Industrial Dataset: Jianbo Zhang¹; Maryam Ghalati¹; Hongbiao Dong¹; ¹University of Leicester
    The Basic Oxygen Furnace (BOF) is crucial in the steelmaking process, converting molten iron into stee by elevating temperature, reducing carbon content, and controlling composition. This stage generates extensive production data, including tabular data and time-series data such as off-gas data and oxygen blowing data. Using domain knowledge, the data was carefully filtered and preprocessed, and 6517 samples were used to train, validate and test. A neural network model was designed to integrate both tabular data and time-series data as input, creating predictive models for end-point temperature and end-point carbon content separately. Several advanced time-series algorithms, including LSTM, GRU, RNNs and CNNs, were tried and compared, with LSTM yielding the best results. After model optimization, the models achieved 93.94% (ą0.02%) accuracy for carbon prediction and 89.03% (ą15°C) accuracy for temperature prediction on testing data, underscoring the significant improvement achieved by incorporating time-series data.

3:00 PM Break

3:30 PM
Fatigue Life and Fatigue Crack Growth Rate Prediction of Additively Manufactured Al 2024 Alloy Using Generative AI and Machine Learning Models: Sneha Jayaganthan¹; R Jayaganthan²; Saurabh Gairola²; ¹Stanford University; ²Indian Institute of Technology Madras
    This study investigates the prediction of fatigue life and fatigue crack growth rate behaviour in additively manufactured Al 2024 alloy using machine learning (ML). The fatigue behaviour of the additively manufactured alloy is influenced by a variety of factors such as defect size and location, columnar microstructure, surface finish, residual stress, etc. Hence, it is difficult to predict the fatigue behaviour using conventional methods. Key test variables, including stress amplitude, build orientation, stress ratio, and post-processing condition, were used as input variables. Experimental data were employed to train various ML models, including Support Vector Machines, Random Forests, Decision Trees, and Convolutional Neural Networks (CNN). Given the limited availability of experimental data, the training dataset was augmented with generative AI models. The pre-processed dataset, comprising both experimental and generative AI data, was utilized to train the above ML models, with hyperparameter tuning performed using kernel functions.

3:50 PM
On the Use of Decoder-Only Transformers to Model Time-Series based Silicon Data: Ricardo Calix¹; Tyamo Okosun¹; Hong Wang²; ¹Purdue University Northwest; ²Oak Ridge National Laboratory
    Silicon prediction in a steel blast furnace is crucial for maintaining high-quality metal production and optimizing furnace operations. In this study, we explore the use of decoder-only transformer neural networks (GPT-like models) for modeling time-series silicon data collected from a Midwestern steel blast furnace. While transformers are traditionally associated with text processing, we adapt the decoder architecture to directly ingest numerical tabular time-series data. Our experiments involve training a GPT model from scratch on silicon data, implementing various data preprocessing techniques, and evaluating different architectural modifications. We find that the model exhibits learning capability and, in some cases, produces promising predictions. However, further refinements are needed to enhance consistency and accuracy. We discuss the advantages of decoder-only transformers for time-series forecasting, the challenges encountered, GPU operational conditions, and potential future improvements. Our findings suggest that transformer-based architectures could play a significant role in time-series modeling for industrial applications.

4:10 PM
Estimation of SOC in Electric Vehicle Batteries Using Machine Learning Models: Bhanu Sree Vijayanand¹; Shrishail C Prabhakar¹; Jayaganthan R¹; ¹Indian Institute of Technology Madras
    The estimation of the State of Charge (SOC) of lithium-ion(LFP) batteries is critical for enhancing the performance of electric vehicles (EVs). Conventional SoC estimation methods utilise coloumb counting and SoC vs OCV measurements. In EV applications, the OCV determination is challenging as the battery pack will be in closed circuit condition during drive cycle. It is important to utilise SOC datasets of LFP batteries obtained under various drive cycles and environmental factors, which could enhance the robustness of predicting its ageing behaviour. In the present work, the SOC data of Li-ion battery (LFP) estimated with the simulated actual drive cycles, were used to train Machine learning models such as Support vector Machine (SVM), XG-Boost, CNN, LSTM for predicting the aging behaviour. The comparative study is made on predictive accuracy of these ML models used in the present work. The mechanisms on aging behaviour of Li ion battery is discussed.