About the talk
Who knew watching paint dry could be so exciting?! Materials that we use every day, from the paints on our houses to the glass on our iPhones, are designed in labs through physical experimentation. With a treasure trove of historical materials data meeting modern trends in data science, the materials industry is poised for an AI revolution. However, pulling scikit-learn off the shelf won't always get you great results, because the physical processes underpinning experimentation and production introduce unique challenges that are not well solved by traditional AI and ML techniques.
This talk will review three challenges encountered during Citrine Informatics's years of experience bringing AI to Materials Design, and will offer solutions with broad applicability to the physical sciences and beyond. Learn about how sequential learning can help mitigate “small data”, probabilistic graphical models can express decades of domain knowledge, and how uncertainty propagation is the key to expressing the instability inherent in physical processes.
Alexandra is an AI Product Manager at Citrine Informatics, with 5 years of experience building tools for people who build AI / ML Models. Most recently, she worked as a tech lead for SigOpt, redesigning their core hyperparameter optimization API, and launching a machine learning infrastructure offering, both of which were topics for her previous MLConf talks. Previously, she has worked at Polyvore, Rent the Runway, Facebook, and Duolingo. Alexandra holds a BS in computer science from Carnegie Mellon University. Additionally, she is the co-organizer for the Bay Area chapter of Women in Machine Learning and Data Science.View the profile
I'd like to introduce Alexandra Johnson to the AI product manager at true sistren, infomatics and welcome. Thank you, girl. So I want everyone to think of a phrase that describes something that is really, really boring. Watching paint dry, comes to mind. So does washing concrete cure? No, I hope you won't use any of these phrases to describe my talk, but these are things that we often used to describe stuff that we don't find very interesting and that we'd rather not you.
But what if I told you that in 2019, the global market was worth 146 million dollars. And the global market for concrete and cement was 440 billion dollars. These are Big Industries and the next generation of materials of these industries design, like the paint for your house or the glass for your iPhone is going to have an impact on our environment, our economy and the way we live our lives. And lucky for us. These industries are just waking up to the potential of artificial intelligence to design their impactful new materials, and that is exactly what I'd like
to talk to you about today. I'm Alexandra. I'm an AI product manager at satrina virmax. Answer train. We're working on building an AR platform for Material Sciences. I came to this field by way of computer science and machine learning. Not my BS, in computer science at Carnegie Mellon. No, I want to take you back to a simpler time, 2015, back when we could do things that today's you incredibly crazy like. In 2015, and I attempted to host a cocktail party and which we were going to use AI
to generate cocktail recipes. I went to a fancy liquor store in Downtown San Francisco and learned about, you know, what are the components of a great cocktail sugar and water? So, I got tequila lime juice, simple syrup. Next, we have to keep your artificial intelligence 11.50 to 1 oz of lime juice and then we started suggesting. Turns out a one-to-one ratio of lime juice, tequila does not make your cocktail disgusting, things like okay, maybe we can express. This is a ratio of
margarita recipes, online. Are the requisition algorithm and we didn't know why we're unable to use case. Fast forward 2020 when I joined the Trine and I am surrounded by folks with chemicals and materials. Cheese. Exactly how to express a problem like this. Turns out she is very similar to creating a high-performing team ingredients ratios and quantities. Anatomical terms. This is called a formulation. What's the Trina Dinosaur? Train have developed a way to express formulations as part of their
platform. All of a sudden there was a generic. Because the train had developed a language for expressing formulation. As a product manager, this is my favorite part. The user interface to the users who are going to be creating new formulations. After learning about this, I was hooked. This huge potential for vertical. I say, I for solving specific problems. That's what I really want to talk to you about the potential solutions to design problems. And the Restless talk, I'm going to walk you through
a few challenges encountered during citrine's years of experience. For me, a ice materials, design service mean, by my generous and intelligent and kind coworkers. This is the start of a typical materials project. Is a handwritten lab notebook? If you're lucky the Datacenter ties, but I need something different. Baby doll, house Excel, but generally you're working with something that's highly unstructured and I mention this because we'll be on our way. Excel equal or good Data Solutions.
Materials data, solution Chiefs, all of the things that scientists love about lab notebooks. Bob Middlebrooks are super flexible. If you want to just do whatever you want to do with these lock notebooks. And that is really working for the scientist. So, whether it's being able to enter whatever. Live notebooks are a little difficult to use because often the second person coming in. And this also comes up in Excel to tell tables as well. You said you measure transparency? What? Wavelength of light is you measure
transparency? I love cooking sour and like a freaking example. Things like that, make a difference in your final product. Last week, we know that any data solution is not going to be an easy start writing it up. Everyone likes structuring data just for the sake of it into this. Besides structuring it for artificial intelligence. We are materials Scientist by the lied to the creation of this material. It is flexible contextualize and incentivize. Why is data is structured? It's time to start searching for that needle in a haystack.
This is a very experimental experimentation for the process. So, generally what happens is someone forms hypothesis about what might help achieve these amazing properties that we want in our new paint and experiment has to be run to verify that we actually can attain the properties that we have. Where should I eat? These experiments. Enter sequential. I'll send you your space efficiently to find the highest performing. Traditional traditional. First, you define a
design space. So, in the cocktail of your alcohol. next greatest, Urban If you, if you don't have any training field. Strategy and this is where you will see a lot of speed up. So there's kind of like a grad student student is the strategy for searching through the space which unfortunately Strategies that can help you search for space more efficiently. Like, Next, you should like the experimental cat has to run the operating the data from their past experience.
And with a well-chosen surrogate model and optimization strategy. Literally parched for data. It was not uncommon for material projects to have only something like We're just talking about how expensive these experiments are no point being. We are so when doing cheer. and speaking of deep neural Nets, so we don't have to do that, but we do have Equation is an equation that describes kind of. if you provide them with, or are you just like acid chemistry underground to just tell you?
Really wants me to make this point we talked about graphical models here. We're talking about a model graph. So, what our solution is issuing code, domain knowledge, as analytic model? All the inputs of the situation may come directly from the data where they might be predicted properties. We're done. By including this domain knowledge into the monograph. We don't waste data, relearning, a relationship that a scientist can just tell us. Senior scientist, scientist. Our users are a scientist and chemist
up until now. This has been, this experimentation is Discovery process. Has been It's often their intuition and they're now. Isn't our goal to to replace them. It's not our goal to to just tell them what experiments are Run. Next. You now just run all them in the process. Like we did with domain knowledge immigration during our expert in the Canada selection process with uncertainty quantification. All physical processes are inherently unstable. Even experts were making up millions of gallons of paint every year. And
also our model is not going to be highly accurate. As we said words. Uncertainty quantification exposes all of this to user instead of just saying, what can I do? I'm here and it's going to be somewhere in the hundreds. We allow our user to get involved in the process. Where are we? Where we, where is our model? Kind of like thrashing around a little bit or user has everything at their fingertips to make data-driven R&D decisions and to say no or not because I want to be able to build models that perform better here. Or I'm ready to start explaining. I want to get you so I prefer it.
Is it a scientist becomes human spider between exploration and exploitation. Is a quote from a question to ask yourself or scientist is designed to support. Exploration learning and intern enrichment. Other expertise. And are we doing that? This talk started out with the impact of materials and chemicals, and opportunities for a i and ended with some possible solutions for common challenges. One might encounter. Experimentation volunteer and development.
Jada is scarce and smart use of Apple stock shares, such as encoding. User is a dumbass. I hope you walk away with some ideas to use in your own work and that you're more inspired to learn about materials and chemicals. Resources that I'll be crying on the website, and if you are super inside the science software, engineering management, research, engineering and talk to you so much for having me and I can take a few questions. Thank you, Alexandra. Are there questions for run for Alexandra?
If there are no questions, this is the last speaker on this track. We might everybody to attend or move over to track one for the startup showcase, which starts at 4:15 on Alexandra. Thank you so much. It was very interesting. Pizza Hut near me.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.