-
Video
-
Table of contents
-
Video


- Description
- Transcript
- Discussion
About the talk
Designing new materials is vital to address pressing societal challenges in health, energy and sustainability. However, the space of hypothetical materials to be considered is incredibly large, and only a small fraction of possible compounds can ever be tested experimentally.
Computational techniques, in particular, atomistic simulation and machine learning, offer an avenue to rapidly invent new materials and navigate this enormous space. Together, they can be used to infer complex design principles and identify high-quality candidates for experimental validation more rapidly and efficiently than trial-and-error experimentation.
In this talk, MIT Professor Rafael Gomez-Bomabrelli discusses how machine learning tools combined with physics-based simulation has enabled materials innovation, and highlights how hyperparameter optimization can make or break a materials design pipeline.
About speaker
Rafael Gomez-Bombarelli (Rafa) is the Jeffrey Cheah Assistant Professor at MIT’s Department of Materials Science and Engineering since 2018. His works aims to fuse machine learning and atomistic simulations for designing materials and their transformations. The Gomez-Bombarelli group works across molecular, crystalline and polymer matter, combining novel computational tools in optimization, inverse design, surrogate modeling and active learning with simulation approaches like quantum chemistry and molecular dynamics. Through collaborations at MIT and beyond, they develop new practical materials such as therapeutic peptides, organic electronics for displays, electrolytes for batteries, or inorganic oxides for sustainably catalysis. Rafa received BS, MS, and PhD (2011) degrees in chemistry from Universidad de Salamanca (Spain), followed by postdoctoral work at Heriot-Watt (UK) and Harvard Universities, and a stint in industry at Kyulux North America.
View the profileThanks so much for having me. My name is assistant professor of Materials Science at the department of today's, going to be a sentinel. I don't think I need to make a basic statement that we need new material. We need them past 20 years. Since I'm out here, we can set this ugly Mexican to the market and you know, Healthcare. Do you see a piece for a computer to get to take Coke machine operation? Computers have increasingly outperformed, people invent new material using
computers. What is the Tipping Point? Where they were going to arrest person? Ever, or we just want to come over ties, past used to be really hard, really expensive, and they become instantaneous. And I'm just very efficient to do your computer. Computers, discontinue oscillator is a Continuum between the first principle simulation that. I have a great day at the races that we know to be happening. How do you spell machine learning and they blackballed Me performance? Of course, Alpha polis
date. What's beside me in this Continuum, right? In this line, that connects the first principles to the machine learning and trying to mix them up in the most effective way. So, please don't feed off of each other and we increase the machine learning and will review the data need of machine learning on the first way that this applies to us is in when we go into a representation learning. So, the first example of Peace in space. How do we improve matters into my own
property relationship material? The structure of a molecule, which is learning. What is the type of Transformations that we need to apply to a place where we can try to do representation learning over the place. And years ago where they moved to know that I really need to learn all the way up to full images. We can do the same picture that Dusty's a format of these days is siggraph formation. Learn how to represent presentation based on information. I'm most of
their second comes from from local environment. It is possible to calculate what color a molecule is going to be that important to me. Like you do diagnostic. Between two successive colors in the rainbow. Outcome of the physical stimulation from the 100,000 times cheaper than actually making them. We've managed to get to about 7 and integrating a representation, in a single stack. I would have let you know that I'm coming here. Flagging pieces together on this happens every time I
let the machine so you can see the first experiment with Ubisoft random met before but I was worse than this Siri that's like two and a half Colors of the Rainbow clothes and then get us to display 5:30, which is the same as a space simulation. What settings would you do about all these things have been really in a standardized, a huge Community pushing? This guy just see whether we can disentangle, right? The occupation of choices were weaned away from Define Union, finally makes
a paper for my Samsung. Deceased or biological set up YouTube into the cell travels to be carried into the cell. Systemically, 25 covered. They are delivering. So this is a design problem. He needs to explore the combinations of amino acids is if we're allowed to use additional. Can you please with no case? We are the most performing. So, what we do for that class of representation learning algorithms. Are we Leverage The sequential nature of the peptide with a one-on-one basis?
So we just used transmission on a presentation in the sequence individual into a speaker. Representation for the what are the settings that make them work program to the top-right. Everything was better than the best thing to do 14. We predicted to get this close and 12 work better than the best husband ever made, you just an example. Midwest, Scooters, Arundel forest, and let you know. We were supposed to see the state-of-the-art and to provide better than any other modem setup. YouTube peptides could be for this.
If this is something that is a generation of that, I just described and how it's going to be presented at a workshop in a few weeks and a life where we work with Lakeland representation. And will you give me for my dick? Monument description last in deep learning representation? Macklemore, some of the biggest things out there when you get food poisoning and we were able to get very good at property place and we were able to get good at attribution error, which part of the
are responsible for the toxicity of what you think is the reason why we only met his answer was the only decent human animals. Building block is extraneous an unknown. Because we were the first people to do this, if this particular application, you can see that The Weeknd Discrete hopes in performance that we moved the needle from South East. Baseline. We know what we're doing way much better performance and answer the actual questions that we care about
something. So everything is booked so far as cleaning price. We make a 4-1 mobile. That is accurate, a lie. And then we're supposed to be raining in Hartland. What is for the property weekend about? That's great. But that's what I set out to do the opposite. I will want to come here to sign for a Target property. So, keeping my desire property. What is the material. What is the? Arrangement of matter and will give me that this was an inspired, unparalleled by General technology machine learning,
who is to sample members of a complex distribution. Now, we want a sample from the distribution of molecules Define emotive listed depressed color. Teal blue look like I will take a molecule we encoded, as we get this customer and now it's a gradient Descent of a property prediction. We can sample some Davis tuition of random and learn to decode. All of these molecules biking. Today, variational autoencoder a picture to make us through the meantime, learnspeed embedding
every Leyton Orient lens to generalize. Adele Rumor, Has it been made yet. Just like the celebrity faces, remind me to know what Layton Pointe map to which property with lots of rules for that allows us to make sure we're sampling from the correct this situation and just focus on finding the right properties through station. Did well. But even back in the first examples of aspirin on the other side Line, TV started looking for a molecule. Now, we win the same money. We can get to pick, which one is this like me better or for the test that we were
after? Alaska. Premier interested in how does it look for stolen item molecules and police are not the same thing. Generates are made from Imperial Point. Infinity repeating three-dimensional pattern. Cordyceps fixed by a prototype on an assignment is for this fine, even though I haven't written the size of a m, b, m c, i don't need to. I just need a way to sample. A pretty picture. But not all the way to just making calls. I know something we don't hear is autoregressive model
example, which atom is going to be on everything sides of the elements. This is how we do. It should be a disciple. Optimize the property. We want it. So we still need a place. We want to minimize the way their main solid high temperatures to satisfy. That is fine. Slow energy. Classic example for business. If anybody in Mobile, where every site interact and learn to sample the right distribution of estate, easy mobile, right? That you don't you set up a temperature and they produced a response to the right temperature
and current examples of different Stice mobile. Buddy, 6:30 a.m. Temperature between order and disorder. Are you can see the properties of volume is not working out for you first. We can also see in that little Plaza with a collar. A nice. And that T-Mobile has to predict their faces with the same company. So we can go back to classical Latin in a, in a Equal samples across a temperature in Madison, Wisconsin, you can see the bottom floor that is as likely rather than more than learned. What time the same likelihood, that same energy rotated frames. Are
you get? This is a task that before we have no idea where they would be able to talk to my sister and I am pleased, that was wasn't, I told her I was completely useless and then the three critical changes in the act. I will say and I hope you don't go faster. Another example to you that machine learning on the physical science into play to invent no more cereal into play architecture and building Tricia about how many layers message passing. Do. I do error message? No message decisions that are
understanding for them under the Patriot Place. Send them as soon as I can and keep asking the difference between being the worst than the red line to architectural being the state-of-the-art by just fine-tuning. The
Buy this talk
Full access
Interested in topic “Education, Training and EdTech”?
You might be interested in videos from this event
Similar talks
Buy this video
Conference Cast
With ConferenceCast.tv, you get access to our library of the world's best conference talks.
