Seminar Talk
Abstract
At a time when Quantum technologies are evolving from laboratory demonstrations into deployable systems and Artificial Intelligence has matured into tools which can turn high-level intent into computational simulations, data analysis pipelines and design workflows, I ask the question ”What can Artificial Intelligence do for Quantum Systems?” and try to explore it from three angles. To begin with, I explore the universal function approximation capabilities of Artificial Neural Networks (ANNs) and leverage them to address a problem that confronts nearly all quantum technologies, namely the measurement and computational burden of quantum state characterization. By learning a map from a limited number of measurements to the entanglement negativity of bipartite ququart systems, I show how one can bypass the computationally expensive iterative reconstruction of the density matrix, while providing reasonably accurate estimates, with 50% fewer measurements and a speedup of three orders of magnitude, made possible due to neural inference. Following this, I explore whether ANNs can detect exploitable structure in sequences generated by quantum random number generators (QRNGs), linear congruential pseudorandom generators(PRNGs) and cryptographically secure pseudorandom generators (CS-PRNGs). As this problem is essentially the ”kryptonite” of pattern recognizers, I explore the predictability using 15 different ANNs and show that the resultant statistical analysis posits this novel framework as a complementary assessment to the NIST SP 800-22 suite, by offering a different perspective on randomness and predictability. Following the application of ANNs to quantum systems, in the second part of the talk, I turn to their latest incarnation as Large Language Models (LLMs), whose emergence has reshaped the scope and practice of modern artificial intelligence. I begin by benchmarking the capabilities of 15 LLMs from 5 different providers on 4 different categories of quantum mechanics problem solving and show that while LLMs excel at symbolic reasoning tasks like derivations, constrained optimization based creative tasks, they struggle with numerical problems. Learning from these insights, I developed a multi-agent AI system Anubuddhi which, by means of a three-layered cognitive architecture, is capable of designing and simulating quantum optics experiments by choosing and arranging the right configuration from a toolbox of optical elements. While this marks a significant improvement over previous automated experiment design approaches by overcoming the dependence on non intuitive intermediate representations and also offers more detailed designs, by choosing not just the optical elements, but also specifying their parameters, sometimes, these can be off by orders of magnitude and I will explain why this happens. Having demonstrated the effectiveness of AI methods for quantum systems, in the concluding part of the talk, I explore the question ”What are the fundamental limits of mechanized reasoning?” by revisiting classical undecidability results from a fresh, Quantum-Inspired Constructive perspective , where I explore how these classical results change when one restricts oneself to finite precision numbers.