By Dave DeFusco
At a time when quantum computing is often described as the future of technology, one student at the Katz School of Science and Health is asking a deceptively simple question: what if we鈥檝e been building quantum programs the wrong way all along?
Prathmesh Joshi, a student in the M.S. in Artificial Intelligence, is developing an approach that brings artificial intelligence and quantum hardware into closer conversation. His research, 鈥淎utonomous, Hardware-Adaptive Quantum Circuit Synthesis via Agentic AI,鈥 aims to make quantum computing more efficient and more practical. To understand the problem, Joshi offers a relatable analogy.
鈥淩ight now, it鈥檚 like designing a car engine first and then testing it on different roads,鈥 he said. 鈥淵ou drive it uphill, downhill and on rough terrain, but the engine wasn鈥檛 built specifically for those conditions. That鈥檚 how quantum circuits are designed today.鈥
Quantum circuits are the instructions that tell a quantum computer what to do, but today鈥檚 systems are extremely fragile. Even small increases in complexity can introduce errors, causing computations to fail.
鈥淐urrent quantum hardware is very noise-prone,鈥 said Joshi. 鈥淓ven adding just a few steps can dramatically increase error. At some point, all you get back is noise, not useful information.鈥
Instead of designing circuits first and worrying about hardware later, Joshi is flipping the process. His system uses artificial intelligence to design circuits with the hardware in mind from the beginning. At the center of his work is what he calls an 鈥渁gentic AI,鈥 a system that learns by interacting with its environment. In this case, the environment is a quantum computer itself.
鈥淐onventional methods generate circuits first and then try to run them,鈥 said Joshi. 鈥淲hat my system does is run experiments in a loop. It observes how the hardware behaves and then adjusts the circuit to better fit it.鈥
This system creates a continuous feedback cycle. The AI proposes a circuit, tests it either on a simulator or real hardware and then improves it based on the results. Over time, the system learns which designs work best for specific machines. The goal is to reduce three key problems: circuit depth, or how many steps are involved, error rates and the extra work required to translate circuits to run on real machines.
鈥淭he whole idea is simple,鈥 said Joshi. 鈥淭he less the circuit depth, the less the error and the higher the chance the algorithm actually succeeds.鈥
One of the most innovative aspects of his research is how the AI adapts to the specific characteristics of different quantum devices. Each machine has its own quirks鈥攈ow qubits are connected, how long operations take and how often errors occur. By feeding this information into the AI, Joshi鈥檚 system learns to tailor circuits to each device.
鈥淚nstead of making one design and running it everywhere, we try to optimize for the hardware we actually have,鈥 he said.
That hardware-focused mindset is grounded in a deeper understanding of physics鈥攕omething Joshi credits in part to Physics Professor Fredy Zypman who has been instrumental in shaping his thinking.
鈥淧rofessor Zypman really helped me understand quantum mechanics at a deeper level,鈥 said Joshi. 鈥淭hat foundation is what makes this kind of work possible.鈥
Joshi is testing his approach using quantum systems developed by IBM, running experiments on problems like optimization and molecular simulation. Early results have already revealed surprises.
鈥淚n one case, we found an algorithm succeeded even when theory suggested it shouldn鈥檛,鈥 said Joshi. 鈥淭hat was very exciting. It shows there鈥檚 still a lot we don鈥檛 fully understand and that experimentation matters.鈥
The research is being conducted under the guidance of Physics Professor Emil Prodan, who sees this work as part of a broader shift in how quantum computing is approached.
鈥淧rathmesh鈥檚 work is important because it connects theory directly with hardware reality,鈥 said Prodan. 鈥淨uantum computing cannot advance on abstract ideas alone. We need methods that account for the imperfections of real machines. This kind of AI-driven co-design is a promising step in that direction.鈥
The project also tackles a major challenge: limited data. Because quantum systems are still developing, there aren鈥檛 many opportunities to run large-scale experiments.
鈥淭here鈥檚 just not enough data yet,鈥 said Joshi. 鈥淲e hit the 鈥榥oise wall鈥 very quickly. That makes it harder for the AI to learn patterns. A big part of the work is finding smarter ways to learn from limited information.鈥
Support from IBM is helping address that challenge. Joshi recently received a $10,000 master鈥檚 fellowship from the company, which will allow him to run more experiments on real quantum hardware.
鈥淓ven a few minutes on real hardware is valuable,鈥 he said. 鈥淚t can lead to meaningful results and even publications. Getting that support from IBM felt like validation that this research is moving in the right direction.鈥
