Recording and questions for Taylor Sparks, Materials informatics: Moving beyond screening via generative machine learning models


Taylor D. Sparks, Associate Professor of Materials Science & Engineering, University of Utah


Friday March 25th, 10am (USA/Pacific)


Technology progresses only as fast as the development of new, advanced materials. Materials discovery has never been more important, but it is far too slow and expensive. Materials informatics has accelerated materials development, but primarily allows us to screen known materials as opposed to truly discover new materials. Here, I will describe our efforts to generate new periodic crystalline materials by predicting crystallographic information file data using generative adversarial networks in conjunction with the newly published DiSCoVeR algorithm that combines a chemical distance metric, density-aware dimensionality reduction, clustering, and a regression model.


A recording of this seminar is available at this link.


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.

Questions answered live

Question Asker Name
Since diamond and graphite consist of pure carbon, can properties of these materials (and many others) predicted only from a chemical composition without a crystal structure? Vitaliy Kurlin
Great presentation! What ways for causal interpretation of these generative results have you attempted so far? Jakob Zeitler
Thanks for the very interesting talk and nice work. In the Compute Gradient step, when you’re generating the vectors on the fly, is there something which ensures that the new vector is chemically feasible? Keith Task
(Bala) The CIF generator is trained to generate CIF cards that are likely to be found close to covex hull. Does it (or is there a way to) accout for synthesizability? Balaranjan Selvaratnam
Any crystal structure can be represented by infinitely many different CIFs. Does this infinite ambiguity make predicting materials properties more challenging than really needed? Vitaliy Kurlin
How can these models esp. GAN or CNN or any other ML model could be used for small datasets (around 1000 to 2000 datapoints) with hundreds of features for material discovery ? Anonymous Attendee
After predicting new materials, there are many experimental parameters involved in synthesizing the materials actually. Can you also predict such synthesize conditions?
Do you know any high throughput experimental material synthesize approach (either in academia or industry) that can explore synthesized new materials (maybe the measured data on new materials can be used in reinforcement learning to improve your models for instance)?
Amir Ataei
Can natural language processing (NLP) help improve the cif prediction? Some contents in cif are texts. lei zhang
Have you compared the Descover-Active learning cycle to Bayesian optimization with Gaussian process model? Balaranjan Selvaratnam
Thanks for a great talk! Short of the lack of an extensive database, can you enivison the application of the GAN structure to investigate a property in amorphous materials? Perhaps using the short-range atomic clusters that are characteristic to a certain composition Karen Ehrhardt
how good is the self checking method, for if the combination of elements is stable or not. Anonymous Attendee
Are the ML modules available for R platform? Novana Hutasoit
1- Does this tool is to work for polymeric materials? and How?
2- Can we directly get the materials’ properties from its 2D XRD pattren?
Ahmed Dawelbeit
Have there been any material predicted that has industrial application Anonymous Attendee
Can you talk about how your approaches to “physical ly motivate” ML compare with methods that embed physical constraints into the operatons themselves such as Equivariant networks. Shyam Dwaraknath
Since a lot of materials gain interesting properties due to their defects (vacancies, for example), can you modify your algorithm to generate different kinds of defected or doped materials? Danish Pannu
How about hybrid materials, like MOFs or hybridg perovskites? Can you featurize separatly the inorganic part of the structure? Gergely Juhasz
Thanks for the insightful talk. Can we incorporate additive manufacturing process parameters and materials properties to tune or discover new material/structure? Novana Hutasoit
This paper ( formally proves that any dimensionality reduction (including t-SNE and UMAP) is either discontinuous (making close points distant) or collapses unbounded sets into a single point (making very distant points identical). Vitaliy Kurlin

From the above list, I removed duplicate questions or questions that were not directly relevant to the talk.

Finally, for everyone who is joining this forum for the first time, welcome! Thanks again for everyone who chose to join us today and who asked questions.

1 Like