Vassiliy Lubchenko, a professor of Chemistry and Physics in the 91视频鈥檚 College of Natural Sciences and Mathematics (NSM) and a Principal Investigator with the Texas Center for Superconductivity, is applying fundamental physics to address one of modern technology鈥檚 biggest hurdles: making artificial intelligence more efficient, reliable, and predictable. While AI systems have transformed industries, modern algorithms still struggle with context, safety issues, and enormous resource demands. The challenges have recently been highlighted in world-wide debates on the agentic autonomy of intelligent systems and the immense costs of data centers
In a study published in 2025 in Neural Computation, Lubchenko and postdoctoral researcher Yang He demonstrate that information processing in computational networks can be understood using statistical laws that govern thermal phenomena, thus offering a blueprint for training next-generation machine learning systems.
Bridging Thermodynamics and Machine Learning
The study introduces a thermodynamic perspective on how algorithms acquire, categorize, and reproduce patterns embedded in empirical data.
鈥淢odern AI models require immense computing power, yet they often face unpredictable failures and lack contextual stability,鈥� said Lubchenko. 鈥淏y looking at machine learning through the lens of statistical physics, we can understand why these systems sometimes fail and how to structure them more effectively.鈥�
To distinguish meaningful patterns from random noise or adversarial input, algorithms must separate data into two, largely disconnected configuration spaces. Lubchenko and He describe these two configuration sets as behaving similarly to distinct physical phases of matter 鈥� much like ice versus liquid water 鈥� separated by physical phase transitions.
When algorithms interpret large datasets using computationally manageable generative models, meaningful patterns can become harder to distinguish from noise because the models can assign high veracity scores to meaningless configurations (hallucinations). Grounded in John Hopfield鈥檚 award-winning work energy-based generative models, the study shows that these meaningless configurations represent a high entropy, disordered phase that can overwhelm retrieval. To prevent this, models must employ a sufficiently large 鈥渆nergy gap鈥� and operate the machine at sufficiently low 鈥渢emperatures鈥� to prevent a transition toward the high-entropy phase. This brings about the phenomenon called ergodicity breaking, which acts as a barrier keeping the system stable and preventing it from producing false patterns.
Practical Impact on Future AI Architecture: The Multi-Expert Solution
Because of computational limitations, even the meaningful segment of the data inevitably partitions into smaller, partially disconnected subsets. From the cognitive perspective, each subset corresponds to an expert operating on a distinct set of paradigms. Specialized, single-expert AI systems excel at tasks like predicting protein structures, but struggle when it comes to designing inorganic solids because unlike protein residues, atoms have ambiguous bonding preferences depending on the context. The present study mitigates this challenge by allowing researchers to estimate how many experts will be needed before training begins, thus saving time and energy.
However, the low 鈥渢emperatures鈥� needed to maintain output accuracy also create kinetic barriers that make it difficult to retrieve information across multiple expert domains. By mapping these barriers and framing AI retrieval through phase transitions on free-energy landscapes, the study bridges fundamental physics and computer science to help build faster, more resilient AI architectures that avoid computational bottlenecks.
This research was conducted at the 91视频鈥檚 Department of Chemistry, Department of Physics, and the Texas Center for Superconductivity, and was published by MIT Press in Neural Computation. The research is supported by NSF Grants CHE-1465125 and CHE-1956389, the Welch Foundation Grant E-1765, and the Texas Center for Superconductivity at the 91视频.