|
Lecturer(s)
|
|
|
|
Course content
|
Week 1: Generation of new molecular structures - Practical implementation of a generative model (Variational autoencoder) for designing new, valid molecular sequences (SMILES) from latent space. Week 2: Design of new high entropy alloys (HEA) - Use of generative models (GAN, diffusion models) to design new HEA compositions and crystal structures. Comparison of outputs and suitability of both approaches for material design. Week 3: Strategies for Efficient Experimentation - Implementing an active learning loop to intelligently select the most informative simulation or experiment to accelerate materials discovery with minimal data. Week 4: Material Properties Optimization - Practical deployment of Bayesian optimization to efficiently find the optimum (e.g., maximum hardness) in a complex design space with the fewest expensive simulations. Week 5: Project: Inverse design of high hardness HEA - Linking a generative model for designing candidate HEA alloys and Bayesian optimization to efficiently find the composition with the highest predicted hardness. Week 6: Simulating physical processes in materials - Building a physically informed neural network (PINN) that respects the laws of physics during training, using the example of solving a 1D diffusion equation. Week 7: Modelling hydrogen diffusion in steel - Application of PINN to model the time evolution of hydrogen concentration in a metal plate. Solving the inverse problem: estimation of the diffusion coefficient from limited experimental data. Week 8: Property Prediction with Limited Data - Application of transfer learning using an advanced graph network (MEGNet) pre-trained on a large database to predict a new property on a significantly smaller, specific dataset. Week 9: From Prediction to Design Principles - Interpretation of a graph neural network model for predicting the phase stability of alloys. Analysis of the effect of individual elements on stability and formulation of hypotheses for the design of new materials. Week 10: Optimization of 3D Printing Parameters - Creation of a surrogate model that maps process parameters (laser power, speed) to the resulting part quality (e.g. porosity) to enable efficient process optimization. Week 11: Automated Experimental Workflow Design - Conceptual design of a closed-loop 3D printing optimization workflow that integrates a surrogate model and Bayesian optimization to autonomously design additional experiments. Week 12: Final Project (Part 1): Design of a New HEA Alloy - Begin work on a comprehensive research design combining multiple learned methods to design a new HEA alloy for additive manufacturing with hydrogen resistance. Week 13: Final Project (Part 2): Completion and Evaluation - Finalise the research design and conduct a peer review focused on the scientific and technical feasibility of the proposed solution.
|
|
Learning activities and teaching methods
|
- Contact hours
- 52 hours per semester
- Individual project (40)
- 40 hours per semester
- unspecified
- 3 hours per semester
- Preparation for an examination (30-60)
- 18 hours per semester
- Presentation preparation (report) (1-10)
- 7 hours per semester
|
| prerequisite |
|---|
| Knowledge |
|---|
| Successful completion of the course KMM/AIM1, knowledge of the basic principles of machine learning applied to material data. |
| Skills |
|---|
| Advanced algorithmization skills (Python) and knowledge of deep learning libraries (PyTorch/TensorFlow), necessary for implementing generative and physics-informed models. |
| Competences |
|---|
| Ability to plan and manage one's own learning and project work. |
| learning outcomes |
|---|
| Knowledge |
|---|
| Explains the principles of inverse material design, i.e., the process of generating chemical composition and structure (e.g., for HEA alloys) based on desired target properties using generative models (VAE, GAN, diffusion models). Understand the theory of physically informed neural networks (PINNs) and how incorporating physical laws (e.g., Fick's laws of diffusion) into the model can replace or refine classical numerical simulations of processes in materials. Describes strategies for effective experimentation, specifically how to use active learning and Bayesian optimization to minimize the number of expensive physical tests or simulations when searching for the optimum (e.g., maximum hardness). Knows the possibilities of predicting properties when data is lacking, including the use of transfer learning from large general databases to specific small material datasets. Is familiar with model interpretation methods (XAI) in order to not only predict a property, but also to reveal which physical or chemical parameters (e.g., the influence of a specific alloying element) are key for a given property. |
| Skills |
|---|
| Design new candidate materials (e.g., high-entropy alloys) by generating their composition and crystal structures from the latent space of generative models. Model the temporal evolution of physical processes, specifically hydrogen diffusion in metals, using physically informed neural networks (PINN) and solve inverse problems (e.g., estimating the diffusion coefficient from limited measurements). Optimize manufacturing process parameters (e.g., laser power and printing speed in additive manufacturing) using surrogate modeling and Bayesian optimization to achieve the desired quality (minimizing porosity). Apply graph neural networks (GNN) for accurate prediction of properties (e.g., phase stability) directly from the atomic structure of the material. Build an automated research workflow that integrates a predictive model and optimization loop for autonomous design of further experiments. |
| Competences |
|---|
| Independently formulate a complex research problem in materials science and design a tailored AI solution for it. Critically evaluate and compare different state-of-the-art AI architectures and justify their selection for a given task. Synthesize knowledge from multiple areas of AI into a coherent research design. |
| teaching methods |
|---|
| Knowledge |
|---|
| Exercises focused on research methodology: Explanation of the principles of accelerated materials discovery (Materials Acceleration Platforms). Analysis of state-of-the-art publications: Critical analysis of current scientific articles on the use of AI in materials engineering. |
| Skills |
|---|
| Project-oriented teaching: Simulation of a real research cycle - from hypothesis generation (material design) to virtual validation. Workshops on advanced tools: Implementation of active learning and Bayesian optimization for experiment control. |
| Competences |
|---|
| Guidance for independent work. Support for teamwork. Development of presentation skills. |
| assessment methods |
|---|
| Knowledge |
|---|
| Exam focusing on the ability to design a suitable AI strategy for solving a complex material problem (e.g., "How would you design an alloy with maximum hardness with a minimum of experiments?"). |
| Skills |
|---|
| Defense of the final research project, where the student presents a newly designed material (e.g., HEA alloy) or optimized process parameter, including validation of the proposed solution. |
| Competences |
|---|
| Ability to work independently and in a team on a complex problem. Critical thinking. Presentation skills. |
|
Recommended literature
|
-
Atwakyire Moses, Ying Gui, Ding Chen. Accelerated material discovery of high-performance Mg alloys via active learning and high throughput multi-objective informed Bayesian optimization. 2025.
-
George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang & Liu Yang. Physics-informed machine learning. 2021.
-
Hao-ran Zhou, Hao Yang, Huai-qian Li, Ying-chun Ma, Sen Yu, Jian Shi, Jing-chang Cheng, Peng Gao, Bo Yu, Zhi-quan Miao & Yan-peng Wei. Advancements in machine learning for material design and process optimization in the field of additive manufacturing. 2024.
-
Joohwi Lee, Ryoji Asahi. Transfer learning for materials informatics using crystal graph convolutional neural network. 2021.
-
Litao Chen, Wentao Zhang, Zhiwei Nie, Shunning Li, Feng Pan. Generative models for inverse design of inorganic solid materials. 2021.
|