The artificial intelligence (AI) machines that guide the world can be divided into three main categories: reasoning machines, learning machines, and discovery machines. Researchers at Washington University in St. Louis are working on some of these rarest machines. New research points to a better way to build discovery machines, thanks to recent work by Shantanu Chakrabarti, Clifford W. Murphy Professor and Associate Dean for Research at Washington University’s McKelvey School of Engineering in St. Louis.
The research, currently published in Nature Communications, builds on previous work on establishing hybrid system architectures, combining “neuromorphic” architectures modeled after the functionality of human neurobiology with systems that leverage quantum mechanics to find optimal solutions to complex problems.
This study shows that these machines can consistently produce state-of-the-art solutions with high reliability and competitive time-to-resolution metrics, Chakrabarty said.
To understand how the new system works, recall these three different types of machines. Reasoning machines are the most common and familiar. For example, ChatGPT acts as a reasoning machine. When you ask a large-scale language model (LLM) to solve a Rubik’s Cube, it is already trained on the exact steps needed for the problem and can provide instructions within seconds.
Now imagine if no one had trained the machine these steps, and the user instead wanted the machine to “learn” all possible steps to solve the cube. That effort requires a learning machine. However, more complex problems require more complex computing, which requires more energy and time. Computer and systems engineers have also gotten better at creating these machines.
In the third category, discovery machines, things get really difficult. Imagine a machine that could not only find all possible solutions to a given puzzle, but also find the fastest and most optimized solution, even with trillions of elements. This type of effort involves harnessing the power of randomness.
In his new research, Chakrabarti essentially provides a recipe for creating AI machines with this kind of capability. These machines are designed to “find the needle in the haystack” and are guaranteed to succeed.
The formula for his team’s discovery machine boils down to a hybrid system of neuromorphic-inspired automatic encoding and Fowler-Nordheim annealing, a tool born from quantum mechanics. “These are the two ingredients you need,” Chakrabarty says. “It’s general enough that it can be applied to any complex problem.”
Autoencoders are techniques for compressing large data streams. With compressed data, the machine performs pattern predictions and repeats the compression process until the predictions are accurate.
Fowler-Nordheim annealing is a method of generating noise and randomness that allows the machine to “tunnel” directly to an optimized solution. This is where researchers found the biggest advantage over traditional computing methods.
New computer chips enable simulated annealing, a method of quantum computing that harnesses the principles of quantum mechanics and can lead researchers more directly to a “Eureka moment,” Chakrabarty said. Using the hybrid system they built, his team can tune the discovery machine to get results.
Teamwork moves the machine
Chakrabartty and his team are participating in this research with collaborators around the world through the Neuromorphic Engineering Institute and through annual brainstorming events such as the Telluride Neuromorphic AI Workshop in Colorado and the Bangalore Neuromorphic Engineering Workshop in Bangalore, India. The study includes co-authors from the Indian Institute of Science, Heidelberg University in Germany, Johns Hopkins University, and the University of California, Santa Cruz. Fayek Ahsan, Chakrabarti’s doctoral student and the study’s lead author, studies the synaptic origins of the process of discovery.
For years, the group has been trying to solve higher-order problems using standardized tests called Ising models. When even neural networks found it too difficult to solve Ising, they came up with the idea of adding a bit of quantum mechanics to enhance the next generation of AI models. But there’s another bonus.
The proposed architecture also differs from previous higher-order Ising models in its convergence guarantees. This means that even if it takes the machine six months or a year to find the answer, they know that at the end of that period they will have the answer. In some cases of supercomputers, researchers could end up waiting a year in vain if they don’t get the prompts right the first time.
Chakrabarty says it’s reminiscent of Deep Thought, the supercomputer in “The Hitchhiker’s Guide to the Galaxy.” We are asked, “What is the answer to life, the universe, and everything?” And it took millions of years for the final answer to be 42, much to the chagrin of its creators. This is not the case with the Discovery Machine, which emerges from the team’s hybrid system.
“This type of machine gives you that guarantee,” he said. “After six months, something useful will come out.”
Ahsan F, Maiti S, Chen Z. et al. Scalability requires only a higher-order neuromorphic Ising machine – an autoencoder and a Fowler-Nordheim annealer. Nat Commune (2026). https://doi.org/10.1038/s41467-026-71937-4
This research was supported in part by research grants (2332166, 2208770, and 2020624) from the National Science Foundation. FA, SC, and AN would like to acknowledge the seed grant from the McDonnell International Scholars Academy and the Memorandum of Understanding between WashU and IISc. Initial research on the XOR-SAT solver was conducted when AN was a visiting scholar at WashU under the Fulbright-Nehru Doctoral Fellowship Program. CST gratefully acknowledges financial support from Pratiksha Trust Grant, India. JS and JK also acknowledge financial support from a Horizon Europe grant (contract number 101147319, EBRAINS 2.0).
