![2
What is materials design?
[2016 Agrawal]
Accelerated Materials Discovery as a Goal-oriented Activity:
? Materials discovery has to be a
goal-oriented activity
? Materials discovery is about
navigating the materials space,
with a purpose](https://image.slidesharecdn.com/arroyavemach2025-250410183224-6f4c03c9/85/Accelerated-Multi-Objective-Alloy-Discovery-through-Efficient-Bayesian-Methods-Application-to-the-FCC-Alloy-Space-2-320.jpg)
![3
Discovery as a (Black
Box) Optimization
3
When queries to the experimental
design space are expensive, we
need to do better than random
exploration
[Brochu]
Model
Policies
Bayesian Optimization:
Prediction of Outcomes +
Prescriptive Policies
Bayesian Methods:](https://image.slidesharecdn.com/arroyavemach2025-250410183224-6f4c03c9/85/Accelerated-Multi-Objective-Alloy-Discovery-through-Efficient-Bayesian-Methods-Application-to-the-FCC-Alloy-Space-3-320.jpg)
![4
Our Current Model Select Single (or many) Experiment(s)
Typical SOA Closed L
[Azimi]
＾Autonomous systems [should] build and
exploit own internal models [and act on
them]￣！N. Freitas
Initial Data
Run Experiment(s)](https://image.slidesharecdn.com/arroyavemach2025-250410183224-6f4c03c9/85/Accelerated-Multi-Objective-Alloy-Discovery-through-Efficient-Bayesian-Methods-Application-to-the-FCC-Alloy-Space-4-320.jpg)
![Autonomous (Materials) Experimentation
? Current studies and methods (applied to materials) have limitations:
? Synthesis is typically limited to vapor/solution chemistry
? Usually optimize to a single objective function
? Based on simple deployment of optimization schemes
[Haase, 2019]](https://image.slidesharecdn.com/arroyavemach2025-250410183224-6f4c03c9/85/Accelerated-Multi-Objective-Alloy-Discovery-through-Efficient-Bayesian-Methods-Application-to-the-FCC-Alloy-Space-5-320.jpg)
![[slide] A Compare-Aggregate Model with Latent Clustering for Answer Selection](https://cdn.slidesharecdn.com/ss_thumbnails/cikm-19-presentationsyoon-191106032817-thumbnail.jpg?width=560&fit=bounds)
More Related Content
Similar to Accelerated Multi-Objective Alloy Discovery through Efficient Bayesian Methods: Application to the FCC Alloy Space (20)
Recently uploaded (20)
Accelerated Multi-Objective Alloy Discovery through Efficient Bayesian Methods: Application to the FCC Alloy Space
- 1. Accelerated Multi-Objective Alloy Discovery through Efficient Bayesian Methods: Application to the FCC Alloy Space Raymundo Arr┏yave, Mrinalini Mulukutla, Trevor Hastings, Ankit Srivastava, James Paramore, Ibrahim Karaman, George Pharr, Brady Butler + all BIRDSHOT Team Thanks to: W911NF-22-2-0106 (HTMDEC-BIRDSHOT program)
- 2. 2 What is materials design? [2016 Agrawal] Accelerated Materials Discovery as a Goal-oriented Activity: ? Materials discovery has to be a goal-oriented activity ? Materials discovery is about navigating the materials space, with a purpose
- 3. 3 Discovery as a (Black Box) Optimization 3 When queries to the experimental design space are expensive, we need to do better than random exploration [Brochu] Model Policies Bayesian Optimization: Prediction of Outcomes + Prescriptive Policies Bayesian Methods:
- 4. 4 Our Current Model Select Single (or many) Experiment(s) Typical SOA Closed L [Azimi] ＾Autonomous systems [should] build and exploit own internal models [and act on them]￣！N. Freitas Initial Data Run Experiment(s)
- 5. Autonomous (Materials) Experimentation ? Current studies and methods (applied to materials) have limitations: ? Synthesis is typically limited to vapor/solution chemistry ? Usually optimize to a single objective function ? Based on simple deployment of optimization schemes [Haase, 2019]
- 6. Discovery Objectives Navigate large FCC HEA Space Discover Pareto Front (UTS/YS, Hardness, Strain Rate Sensitivity) Explore effect of SFE Demonstrate acceleration through a Bayesian Materials Discovery approach 6
- 7. Materials Science & Engineering C High Throughput Materials Discovery for Extreme Conditions (HTMDEC) 7 ? Multiple information sources ? Resource optimization through Bayesian Optimization ? Capability of prescribing optimal queries in batch mode ? Highly efficient synthesis and processing workflows ? SoA high strain rate, high- throughput characterization ? Tight integration between experimental and simulation tasks
- 8. 8
- 9. 9
- 10. Alloy Search and Design ? Compositional Space of Interest: CoCrFeNiVAl ? Alloys sampled at 5 at% resolution (~53k alloys1) ? Use HTP CALPHAD, ML, and physics-based model calculations to identify feasible space ? HTP filtering of alloy space based on constraints - FCC alloys at 700 C2 - Solidification range4 < 100 K - Thermal conductivity3 > 5 W/mK - Density5 < 8.5 g/cm? ? Constrained space6 for medoid selection ? K-Medoids clustering of alloys in feasible space 1 2 3 4 5 Constraint Design for Filtering of Alloy Space 6 Alloy Space White-Meets Constraints Red-Does Not Meet Constraints Single Phase FCC Thermal Conductivity Solidification Range Density Feasible Space: ~4% 10
- 11. Department of Materials Science and Engineering Reduced Design Space https://umap-learn.readthedocs.io/en/latest/index.html https://umap-learn.readthedocs.io/en/latest/how_umap_works.html Compositional Space UMAP Projections K-Medoid Clustering 6nary HEA: Al-V-Cr-Fe-Co-Ni Rapid approach to >10m possible alloys Constrained Sorted by Co Sorted by Entropy Low SFE High SFE Every color corresponds to a k- medoid cluster Design of Experiments (DOE) approach enables space-filling designs over non-convex multi- dimensional spaces Search over two alloy Stacking Fault Energy(SFE) regimes: ＾Low￣ (< 50 mJ/m2): 1,550 alloys ＾High (> 50 mJ/m2): 887 alloys Low High 1
- 12. Materials Science & Engineering C High Throughput Materials Discovery for Extreme Conditions (HTMDEC) 12 26 mm 7 mm Homogenization 35 mm 14 mm Machining & Sample Preparation Bulk Mechanical Property Characterization XRD Tension Microhardness Vacuum Arc Melting Batch Material Production Thermomechanical Processing EDS 45 mm 4 mm Forging Materials Design & Optimization , Iterative Loops Chemical Analysis Materials Design Nanoindentation , LIPIT, Spherical Microindentation Target Property Data fed into the database and the models 16 samples / iteration 8 μm Microstructure & Homogeneity Evaluation SEM/EBSD Microstructural / Mechanical Property Characterization
- 13. 80 different alloys fabricated via Vacuum Arc Melting with high chemical precision Microstructure was characterized. Alloys verified to be FCC single phase. 8 μm Mechanical properties were determined via Vickers hardness tests and isothermal tension experiments 13
- 14. 140 190 240 1.5 2.5 3.5 4.5 5.5 6.5 Modulus (GPa) Hardness (GPa) 0.001 0.006 0.011 Strain Rate Sensitivity Exponent (m) Nano-indentation 14
- 15. Dealing with Multiple Objectives and Constraints
- 16. Materials Science & Engineering C High Throughput Materials Discovery for Extreme Conditions (HTMDEC) 16 Multi-Objective optimization to discover the Pareto front.
- 17. Materials Science & Engineering C High Throughput Materials Discovery for Extreme Conditions (HTMDEC) 17 Key Acceleration: Batch BO
- 18. Department of Materials Science and Engineering High Stacking Fault Energy Regime Pareto Surface Alloy UMAP 1 Results C Materials Discovery
- 19. Department of Materials Science and Engineering ￥ Employed multi-objective Batch BO to maximize 3 objectives: UTS/YS, Hardness at trusted strain rate, Strain Rate Sensitivity (SRS) ￥ 1000 GPs created to map 6D composition space to the objective space ￥ Expected hypervolume improvement employed as the decision-making policy enhanced Pareto front approximation ￥ 5 iterations completed in both low SFE and high SFE spaces, making 8 observation per iteration 1 Results C Materials Discovery
- 20. 20 SFE Effect ?
- 21. Were we lucky? 21 ? In typical ｀academic¨ work on BO, one has the ground truth function that must be optimized ? Hundreds of BO runs are done to examine convergence of an algorithm ? However, when BO is used for materials discovery, we don¨t have a ground truth, nor we have the luxury of running the BO scheme multiple times ? The question, then, is how do we know that BO work? ? Maybe it was pure luck? ? This remains a frontier problem
- 22. 22 An empirical test: Were we lucky? Bayesian methods dominate random search in synthetic materials discovery problems
- 23. 23 A statistical test (comparing against random chance): Were we lucky? ~15 Pareto points in 3-objective space out of 40 draws (for each sub-campaign) Our results suggest that it is < 0.16% likely that we found the Pareto set by luck
- 24. Materials Science & Engineering C High Throughput Materials Discovery for Extreme Conditions (HTMDEC) 24 Accelerated Materials Discovery Copper 5,000 BC Bronze 3,500 BC Iron 1,500 BC Damascus Steel 500 AD Type Metal 1500 AD Aluminum 1825 AD Stainless Steel 1925 AD Ni-Superalloys 1940 AD HEAs 2004 AD 1 month ago 1 year 10 years 100 years 10,000 years 1,000 years parallel random search birdshot random search ? BIRDSHOT discovered, in less than a year, the Pareto front by querying 0.15% of the potential ~50,000 designs ? As a (conservative) comparison, finding the best material by random chance at 50% success rate would take ~ 25,000 queries ? Each alloy takes ~1 month to make and test, so this effort would be equivalent to ~ 2,000 years. ? If one could parallelize this effort, this would be equivalent to ~ 150 years. 100,000 years
- 25. 25 Thanks!