Smaller, faster, smarter: Clean Energy Lab reimagines AI

Researchers working in the Clean Energy Lab at UTSC have developed an architecture for training AI that has the potential to eliminate AI hallucinations, making it more useful for scientific discovery and simultaneously reducing power wastage.  

According to principal investigator Associate Professor Oleksandr Voznyy, a member of the Acceleration Consortium, the lab is seeking a patent on a new approach, an algorithm that may significantly improve the outputs of chemistry-based and medicinal AI applications... and even large language models (LLMs).

AI hallucinations are instances where an AI system produces outputs that sound plausible but are factually wrong, irrelevant or entirely fabricated.  

Standard AI models hallucinate because they cram dozens of different concepts into the same artificial "neuron,” according to Prof. Voznyy.  

“Imagine a student cramming for an exam who accidentally memorizes textbook facts and gossip simultaneously; when asked a question, they mix up the facts with the gossip, producing a plausible-sounding hallucination. Standard AI spreads information across continuous, entangled web-like networks, making it impossible for the model to tell where a true physical fact ends and a random guess begins.”

By contrast, the architecture from the Clean Energy Lab acts like a strict distillation column for information.

“It forces the AI to assign exactly one, pure concept to each artificial neuron—a property called monosemanticity. Because every concept is cleanly isolated into its own independent channel, the AI cannot "smear" facts together or invent connections that do not exist,” said Voznyy. “If a concept is not explicitly supported by the data, its output valve is shut off entirely, physically preventing the model from hallucinating made-up details.”

The benefits extend well beyond stopping hallucinations. This method also solves the infamous "black box" problem of artificial intelligence. “Because every active neuron represents a single, clear concept, we can look inside the AI's brain and see exactly why it made a decision. If the AI exhibits a bias or a specific error, we do not have to retrain the entire model; we can simply turn off the specific neuron responsible, like flipping a light switch,” Voznyy explained.  

In addition to increasing accuracy and accountability, an additional benefit—and one that addresses a significant downside of this technology—will be reducing the energy costs incurred by modern AI.

"Standard models waste enormous amounts of electricity running millions of redundant calculations on noisy data. By aggressively shutting down useless neural pathways during training, our models become up to 100x smaller and faster, allowing frontier-level AI to run efficiently on ordinary laptops or local medical equipment rather than on massive, energy-hungry data centres,” he said.

In recent years, researchers in chemistry have been making innovative use of AI to explore never-before-seen chemical interactions and compounds. By combining machine learning with self-driving labs that can run hundreds of experiments and use the results to steer away from dead ends, they can accelerate the development and analysis of new materials. One example: a team of researchers at U of T Engineering, using AI-driven materials design, has recently discovered a new set of metal alloys that retain their strength under extreme conditions.

Turning to the subject of medicine, Voznyy said: “We are applying this architecture to diagnose complex, difficult-to-treat conditions like Myalgic Encephalomyelitis / Chronic Fatigue Syndrome (ME/CFS) using blood metabolomics, which measures thousands of small molecules in a patient's bloodstream."

"ME/CFS has historically been challenging to diagnose because patient datasets are relatively small and blood chemistry contains significant biological noise. Standard AI gets overwhelmed by this clutter, but our algorithm starves the noise to pinpoint the exact, pure molecular biomarkers that signal the disease.”

Voznyy said researchers can use this architecture to improve the generalization of small and noisy scientific experimental datasets. “We are currently also developing faster and more accurate machine learning potentials (MLIPs) that can simulate atomic systems, bypassing more expensive quantum chemistry methods.”

Asked what had motivated him to work on improving AI, Prof. Voznyy said the work was a response to limitations within the existing tools available for research. "This work grew out of pure frustration with existing AI tools in our lab.”

“Commercial AI is designed for "big data" - millions of internet images or text files. But real-world science does not work that way. In chemistry, datasets are often small, expensive to collect, and full of experimental noise.”

As an example, he mentioned Morgan fingerprints. Sometimes referred to as ‘barcodes for molecules,’ Morgan fingerprints are a method used to convert a chemical structure into a machine-readable numerical representation.

“Standard chemical representations fail completely when you have fewer than 100,000 samples,” Voznyy said. “Because standard AI simply memorizes the noise instead of learning the underlying molecular rules.”

He realized that to make AI work for real science, it was necessary to fix its core training dynamics. “This new architecture directly feeds into all our clean energy research, from designing better solar materials to discovering energy-efficient catalysts. By forcing the AI to ignore dataset noise and discover true, generalizing physical laws from small amounts of data, we can accelerate scientific discovery without needing millions of lab experiments or massive computing clusters.”

“Large language models (LLMs) currently hallucinate because standard training rewards them for sounding fluent, not for being factually precise. They entangle facts, grammar, and tone into the same parameters. In high-stakes fields like medicine, law, or scientific research, even a 5% hallucination rate makes an AI a dangerous liability.”

"Our approach fundamentally alters how LLMs process information by forcing facts and reasoning steps into isolated, monosemantic nodes.”

“This completely transforms AI safety and auditing. Imagine being able to open up an LLM and physically inspect the logic circuit behind its answer. If a model generates a false medical recommendation, engineers can trace the exact node responsible for that claim.”

“This turns LLMs from untrustworthy, unexplainable black boxes into precise, auditable, and highly compressed engines that can be safely trusted in critical applications.”

Reducing noise within AI models will, it is hoped, also reduce computational waste while helping researchers extract reliable insights from limited scientific data.

The result could be smaller, more transparent AI systems that accelerate discovery without requiring massive datasets or energy-intensive computing infrastructure.

Republished from the Department of Chemistry

Thumbnail image: AC member Oleksandr Voznyy, an associate professor at U of T Scarborough, in the Clean Energy Lab. Photo credit: Dylan Toombs

‍