Early tests show ‘AI supermodels’ can speed up materials discovery by 100x (or more) with minimal data

Despite all of the bluster surrounding AI’s transformative promise, some R&D projects—especially in areas like battery materials development and drug discovery—face significant bottlenecks when relying on these tools, often requiring days and sometimes weeks or even months of GPU processing time on enormous datasets. Enthought—among other organizations—are investigating leaner approaches to tackle these limits, which it has dubbed “AI Supermodels,” or “next generation materials informatics.”

The aim? To make better predictions about how, say, a given material or compound will perform in a certain condition. As Mike Connell, chief operating officer of Enthought, explains, “In fact, it’s reasonable to say that the strength of your R&D really depends on how well you can make these kinds of predictions. Faster predictions mean faster R&D. Better predictions lead to better R&D.” Faster and more accurate predictions, in essence, directly translate into faster and better R&D outcomes.

Early tests, described in a recent Enthought webinar on “AI Supermodels”, demonstrate this potential for data-efficient AI in R&D. This is supported by research into data-efficient AI methodologies, such as a 2022 Nature Communications study from Los Alamos National Lab on quantum machine learning, which shows that quantum machine learning models can reliably generalize using only a modest (polynomial or even polylogarithmic) number of training examples relative to the number of qubits and parameters.

Doing more with a lot less

Such approaches can achieve high-fidelity results with dramatically less data. In the webinar, Connell described an example requiring as little as 1/100th of the typical input and potentially reducing measurement requirements from thousands of points to approximately ten. Such techniques can compress fine-tuning and analysis times from days or weeks down to minutes, slashing both compute processing time and overall project calendar time.

Connell describes an example at Los Alamos Labs involving quantum tuning of incredibly sensitive quantum sensors. Imagine meticulously adjusting multiple lasers to pinpoint a single point in space, a process requiring weeks of painstaking manual work by expert researchers. “It normally takes days or even weeks to tune this, and they got it down to four minutes using this model,” he said. “And so that’s a reduction of about 1000 fold, right? 100 fold to 1000 fold. So that’s a huge increase in speed.”

In an example involving x-ray diffraction analysis for materials design shared in the webinar, the challenge was interpreting complex X-ray diffraction patterns to determine the atomic structure of materials. Traditionally, this was a slow, iterative process relying heavily on expert intuition and often taking days or even months. Using an AI supermodel, however, this analysis time was dramatically cut to mere minutes – achieving results comparable to expert intuition in a fraction of the time.

Yet the technique “hasn’t reached the mainstream yet,” says Connell in the webinar. That said, according to Enthought, they see potential for these data-efficient solutions to transform R&D workflows.

Connell also discusses the name “AI supermodels.” “People are figuring out how to use AI to superpose or combine the intuition, theory, and data to build even more powerful models,” he said. “That’s why I call these AI supermodels, because they superpose the different sources of information, and they get superior models.”

On “AI supermodels” and surrogate models

AI supermodels build on the core principle of surrogate models—using simplified approximations to bypass resource-intensive simulations—but integrate more domain knowledge and theoretical constraints to reduce data requirements while boosting predictive fidelity. While both approaches aim to bypass resource-intensive simulations by using simplified approximations, AI supermodels distinguish themselves through key enhancements:

Integration of domain knowledge and theoretical constraints: AI Supermodels, by contrast, go further than traditional surrogate models in that they actively integrate theoretical constraints and domain-specific physics directly into their architecture. That means they don’t just memorize data; they incorporate what Connell calls “the physics we know,” effectively restricting the model’s form in ways that let it predict accurately with far fewer data points.

As Connell explains, AI supermodels “integrate more domain knowledge and theoretical constraints to reduce data requirements while boosting predictive fidelity.” This is in contrast to basic surrogate models which, as Connell notes, are “surrogate in the sense that it doesn’t know any physics. It just is kind of a look up table… generating it from memory, in a sense.” AI Supermodels, however, are designed to incorporate “the physics we know,” allowing them to make more accurate predictions, especially when data is limited or when extrapolating to new conditions.

Data efficiency and predictive fidelity: By embedding domain expertise, AI supermodels require significantly less data to achieve high-fidelity results compared to traditional machine learning methods. This data efficiency is key in R&D settings where data acquisition can be expensive and time-consuming.

To recap, AI supermodels build upon surrogate models, but introduce some improvements in R&D prediction. While surrogate models act as “look-up tables,” AI supermodels integrate theoretical constraints and domain physics. Taking this a step further: Surrogate models are data-driven while AI supermodels combine intuition, theory, and data. In addition, AI supermodels require significantly less data than traditional Deep Learning approaches, reducing measurement points considerably. AI supermodels can also drastically reduce processing time, sometimes from weeks or days to minutes. Finally, AI supermodels offer improved extrapolation and higher fidelity as a result of their theoretical grounding. In essence, AI supermodels move beyond data-driven approximations by incorporating scientific understanding, leading to faster, more data-efficient, and more reliable predictions for R&D.