Speaker
Tonio Buonassisi, Professor of Mechanical Engineering, Massachusetts Institute of Technology
Date
Monday August 15th, 10am (USA/Pacific)
Abstract
Inverse design entails the intentional creation of a material embodying user-defined properties. It’s widely accepted that only a small fraction of possible materials has been synthesized, and thus, an inverse-design tool can help navigate unexplored spaces more resource-efficiently than current approaches. Historically, inverse design was envisioned considering available toolsets, initially heuristics and intuition, later first-principles calculations, and more recently, machine-learning (ML) methods that leverage elements of the latter two as training data or constraints.
In this presentation, I’ll review the state of the art of ML-driven inverse design, and highlight outstanding gaps to be addressed. Highlights include novel materials representations that reflect underlying physics of elements and their diverse combinations, ML algorithms designed to capture and generalize patterns from known compounds, refinement approaches designed to eliminate junk data, and early attempts to integrate inverse design into experimental workflows. I hope the audience will leave energized to contribute to this high-risk, high-reward research area, which in this decade promises to deliver surprising breakthroughs and commensurate industrial impacts.
Recording
A recording of this seminar is available here.
Questions
If you are unable to ask questions live, please feel welcome to ask any questions following the talk here and we will ask the speaker to check afterwards. Whether they will be able to answer questions or not depends on the speaker’s availability.
Answered Live
| # | Question |
|---|---|
| 1 | How to better inversely design devices? That requiest different materials/electrodes/synthesis/interfaces, but are very complicated. Any recommendation on effective ways for devices? |
| 2 | Why did you pick the Autodesk product? Could you not this this with SAS JMP? |
| 8 | Follow up: how does the FCTP lead to the invertibility you were targeting? |
| 10 | Very stimulating talk Tonio (I had to disconnect for dinner but will catch up later!). At this point in time, are limitations in data or model architecture more important to overcome for general inverse design? |
| 12 | The crystal generation, when the high-dimensional space to be optimised doesn’t not have very large variations (compartively flat (low variance) high-dimension space), is very sensitive to the acquisition function of the optimization method. If possible can you give us an insight into what latest kind of optimization approaches you employ? |
| 15 | From your experience, what is thresold for energy convex hull (estimated from DFT computational phase diagram) for which predicted materials might be made? |
| 17 | do u do any work on SiC or battery materials? |
| 18 | Thank you very much for this great presentation. Please which featurizers are based for example on Structure for classification Machine Learning? |
| 19 | Great Talk! Can you please explain why composition models achieved better performance than structure model? What is the training and test dataset? |
| 20 | Are the computational tools can be used? where can we get it? in MIT? |
| 22 | For inverse design of crystalline materials, what is the biggest challenge right now: quantity/quality of data, effective crystalline representation + invertibility, or training VAEs/GANs etc. properly? |
| 23 | May I ask what is the dimensionality of the FTCP? |
| 24 | Is inverse design by the FTCP method interpretable? That is, can we find out why the inversely design materials possess the targeted properties? |
| 25 | In any project description I miss the inclusion of recycling and reuse of the original elements. Is it possible to include that or are the missing data the hurdle. |
| 26 | Are there any good ways for extrapolation to discover unprecended materials? |
Not Answered
| # | Question |
|---|---|
| 21 | Can we able to predict materials properties with different synthesis method how there property will vary through computational method |
| 27 | How close to equilibrium are the structures that the VAE algorithm outputs? |
| 28 | As a professional in the field, what’s your advice for a freshmen student in materials science to boost his performance and to be more productive in this field? |