EmTechMIT, Cambridge, MA: The scientific method, observation, hypothesis, and experiment, has fueled centuries of breakthroughs, but Rafael Gomez-Bombarelli believes it’s time to leave trial-and-error behind. At EmTechMIT, the MIT professor and cofounder of Lila Sciences cast a vision for “scientific super intelligence,” driven by a new fusion of artificial intelligence and experimental science. “Our ambition,” he began, “is to achieve scientific super intelligence to address humankind’s greatest challenges, doing science as the best scientist out there, or better.”

The Scientific Method Reimagined

Historically, science has been slow and manual, with researchers laboring over experiments, iterating on hypotheses, and often spending decades to move an idea from conception to practical impact. Gomez-Bombarelli described this as a “20-year gap from hypothesis to commercial scale,” especially in fields like materials science, where new materials can transform industries but are held back by the costs and inefficiencies of traditional research.

Most experiments, he emphasized, fail. “The valley of death” is littered with promising ideas that prove impossible to scale, optimize, or manufacture. The process is inherently stochastic, part exploration, part luck, which stalls innovation and bottlenecks progress.

How AI Closes the Loop

So how does artificial intelligence radically alter this paradigm? Gomez-Bombarelli explained that today’s machine learning models do much more than parse vast swaths of scientific literature; they also simulate real-world systems and propose creative solutions. “It’s a machine that reads the literature for us,” he said, referencing advances in representation learning and property prediction for materials science.

Beyond literature review, AI now runs simulations at unprecedented scale, generating hypotheses about everything from the atomic structure of materials to reaction pathways for industrial chemistry. Crucially, these models operate without the human intuition of domain experts, they “exploit data and compute better,” identifying patterns invisible to manual inspection.

Self-Play and Scientific Super Intelligence

A paradigm borrowed from game-playing algorithms is now driving a new era in scientific research, self-play. Just as AlphaGo shocked experts by discovering winning moves that no human had ever tried, Gomez-Bombarelli believes AI can similarly “self-play” science, recursively iterating experiments against the rules of nature. “AI models recognize what is the most brilliant next experiment to be run. In our case, we want to create self-play for science. Nature itself tells us, ‘You did this right, you did this wrong.’”

This recursive feedback loop marks a departure from trial-and-error. Instead, experiments are chosen and scored algorithmically, with both successes and failures used to steer the course forward. “Building up a virtuous cycle of scientific self-play,” Gomez-Bombarelli explained, “lets AI propose unexpected hypotheses and carry out experimental testing automatically.”

Automation of the Lab: The AI Science Factory

Perhaps the most radical shift is occurring inside the laboratory itself. Gomez-Bombarelli’s group has built “AI science factories”, automated experimental platforms that close the loop between simulation, hypothesis generation, and real-world validation. These systems can run dozens (or hundreds) of experiments in parallel, rapidly gathering data, optimizing conditions, and validating predictions.

“The first example,” he noted, “was a catalyst scanned for hydrogen production. This system took a standard 96-well plate used in biology and adapted it for materials science, a domain previously lacking standardized lab protocols.” The automated lab pulled together the disparate strands of simulation, experimental validation, and iterative improvement, forming an ecosystem where AI and hardware continuously refine the scientific process.

Tackling Grand Challenges with New Materials

Materials are at the heart of Gomez-Bombarelli’s work, and at the heart of global challenges like climate change and energy sustainability. “Materials are critical for everything, energy, sustainability, health,” he emphasized. “How do we harvest energy from the sun, create new batteries, or enable carbon capture?”

The bottleneck, he observed, isn’t just designing new compounds on a computer, it’s scaling their manufacture and integrating them into real-world devices. Questioned about breakthroughs in energy storage, he referenced technologies from the 1990s like advanced battery chemistries, but argued that “there are many more classes of materials yet to be discovered.” AI-powered automation will, he believes, unlock these materials faster, cheaper, and at a scale previously unattainable.

The Data Engine: Beyond Human Limits

Gomez-Bombarelli stressed that the future isn’t just about running more experiments; it’s about running smarter ones. “Large language models can now access every written word, all the literature, all the internet,” he said, but the missing ingredient is “creative new avatars” of experimentation. He challenged the field to “upgrade the intelligence of our models… set up a virtuous cycle where the AI proposes and optimizes experiments in pursuit of unexpected advances.”

He sees this approach as a foundation for “generalizing” scientific thinking across new fields, from chemistry and life sciences to nanomaterials and structural engineering. “Just like AI can master new games, we believe it will unlock expertise in many areas of science simultaneously.”

Are We There Yet?

When asked about the current status of this revolution, Gomez-Bombarelli was unequivocal: “This is happening. Large language models connected to scientific verification already show sparks of super intelligence in certain areas.” He described the active connection of AI models to both literature databases and lab hardware: “Today, a model can read the literature, run simulations, and interact with a laboratory in the same sequence of events.”

There are, of course, cautionary notes. “We’re taking safety and security very seriously,” especially in “self-play in the laboratory,” which demands strict limits and guidelines. Automation does not mean indiscriminate experimentation, but rather careful orchestration of resources and checks against real-world risk.

The Road Ahead

The end of trial-and-error science could mean a future where discovery accelerates exponentially. Human creativity remains essential, AI does not replace the scientist, but augments and challenges them via new pathways for exploration. “How do we have humans be super-powered by AI?” Gomez-Bombarelli asked rhetorically. “How do we bring AI to propose really cool, unexpected hypotheses?”

For technologists, this session drew a clear roadmap: Combine simulation, automation, and recursive feedback to make scientific inquiry vastly more efficient and ambitious. The bottleneck is no longer the scientist’s stamina or the laboratory’s throughput, it’s the imagination of the next experiment, the intelligence of the next model, and the ability to learn from all experimental data, not just the successes.

The next century of scientific progress, Gomez-Bombarelli suggested, may arrive not through serendipity, but by design. “Unlock generalization across science,” he urged, “and bring the future into the present.”

For more information, please visit the following:

Website: https://www.josephraczynski.com/

Blog: https://JTConsultingMedia.com/

Podcast: https://techsnippetstoday.buzzsprout.com

LinkedIn: https://www.linkedin.com/in/joerazz/

X: https://x.com/joerazz