AI + quantum computing: ĢƵ, NVIDIA, and Pfizer have combined transformer-based generative AI with quantum computing to automatically generate high-quality quantum chemistry circuits more efficiently than traditional optimization methods.
Practical pharma impact: The approach was used to prepare molecular ground states and validated on ĢƵ’s Helios hardware, demonstrating a path toward larger-scale computational chemistry and drug discovery.
Long-term vision: The team aims to build quantum foundation models that learn from increasingly complex quantum data, eventually enabling AI to design circuits for molecules too large for classical simulation.
Quantum computing has long promised a future that expands what we can do with compute — for example, in molecular simulation, materials discovery, or pharmaceuticals development. But between that promise and practical utility sits a stubborn bottleneck: quantum state preparation.
To run any algorithm on a quantum computer, you must first put the qubits in the right starting state. Think of it like setting up a Rube Goldberg machine- except in this case, you’re not sure exactly which initial setup will give you the results you want. This is what makes quantum state preparation so important: your choice of initial state dictates the accuracy and cost of the rest of the calculation.
We teamed up with NVIDIA and Pfizer to tackle this problem, with an eye towards developing meaningful industrial workflows. The result is a new generative quantum AI framework, called ADAPT-GQE, which we consider to be a canonical instance of GenQAI. ADAPT-GQE uses quantum data to train transformer models that ultimately synthesize quantum chemistry circuits faster, with better outcomes, in a sort of ‘virtuous cycle’.
Ultimately, this means we have developed a new interface between quantum computing and AI. By treating quantum circuit generation as a language modelling problem, we now have a system that can generate high-quality ground-state preparation circuits - with comparable or improved state preparation accuracy.
The Magic – and Difficulty – of Computational Chemistry
The goal of computational chemistry is to learn about chemical properties without performing expensive, time-consuming, and sometimes dangerous “wet-lab” experiments.
In principle, you can replace the majority of your physical experiments with computer simulations, saving billions of dollars and years of time.
In reality, computational chemistry is very tricky. To accurately simulate a chemical inside of a computer, you have to build it from the ground up. You start with a collection of atoms (in the case of imipramine, you have 19 Carbon atoms, 24 Hydrogen atoms, and 2 Nitrogen atoms). Then, like Nature’s ‘lego’, you assemble those atoms into a molecule: you set bond lengths, strengths, angles, interactions, and so on.
This is not straightforward: a single molecule can exist in many forms; with different angles, rotations, etc. We will call these different forms ‘conformations’.
Then, to actually estimate chemical properties, or to explore chemical reaction pathways, you have to reproduce the detailed physics that goes on at the atomic level: take your chosen conformation then figure out how each orbital is occupied, how the electrons are interacting with each other or the atomic nuclei, how is the addition of heat or a catalyst going to affect things.... it gets complicated, quickly.
Despite all this, computational chemistry is a powerhouse in pharmaceutical development. Right now, pharmaceutical companies save money and time by simulating as much as they can on computers, avoiding time consuming and expensive laboratory experiments. However, even with ~50 years of development, the existing classical methods have very real limitations.
This is where quantum computing comes in: this new computational paradigm can elide those limitations because it has many of the “hard parts” (like superposition or entanglement) natively encoded. Used correctly, quantum computing promises to break old barriers, further improving margins for pharma companies across the globe while contributing to meaningful, impactful, discoveries.
A Virtuous Cycle: Using Quantum Data to Train AI, Which Then Designs Better Quantum Circuits
While quantum computational chemistry is one of the strongest candidates for near-term quantum advantage, current hardware is still in the earlier stages of development. With limited qubits and error rates, algorithm designers need to make every gate count, keep circuits shallow, and be able to tolerate some level of noise.
This is where generative AI enters the picture.
Instead of hand-designing chemistry circuits and laboriously experimenting to see how well they run, there is another idea: what if we trained an AI to solve the problems that quantum computational chemistry faces?
Using this approach, not only can we save time and resources; but we can shorten the timeline to realize practical results. With better state prep and other circuits, applications that were once considered far in the future come into view.
Our first attempt at this is called ADAPT-GQE. The central idea behind ADAPT-GQE is deceptively simple: instead of laboriously searching for good quantum circuits from scratch, train a transformer model to generate them directly.
Importantly, the framework is model-agnostic, which we showed by deploying it on complementary transformer architectures - (a pretrained LLM) and (trained from scratch).
From Iterative Optimization to Generative Models
The initial goal here is to find the ‘ground state’ of the molecule imipramine (this is the electronic state with the smallest amount of energy stored inside it). To do this, you have to find the right ‘state preparation circuit’, as described above.
Until now, a leading method for finding the ground state with quantum computers was the ‘Variational Quantum Eigensolver (VQE)’, a hybrid quantum-classical approach. The VQE process starts with a ‘guess’ circuit for a particular conformation of the molecule. The quantum computer runs the circuit to measure the associated energy of the molecule. This result is fed back into a classical optimizer that then tweaks the circuit parameters, hopefully resulting in one with a lower molecular energy. This loop repeats until a minimum energy is found.
Unfortunately, VQE has a few severe limitations that make it infeasible for widespread use. The recently proposed ADAPT-VQE was a crucial step forward meant to address some of the issues with “plain” VQE. In ADAPT-VQE, instead of starting with a guess for the initial circuit, the process builds a circuit in steps by selecting operators from a pool(typically using gradient information) and optimizing. This approach can be more effective, but unfortunately still grows too large too quickly.
This is where the joint team jumped in.
Combining the best of all worlds, the team’s new framework, ADAPT-GQE, combines AI with the ADAPT-VQE to create something entirely new – and something that, so far, is a scalable, hardware-validated pathway toward automated quantum circuit synthesis.
First, transformers (in this case, Nemotron and Gemma) are trained via supervised fine-tuning on ADAPT-VQE data. In this way, the old method isn’t thrown away but is instead treated as a high-quality data-producing “oracle”.
Then, once the transformer has been initially trained, it defines a distribution over circuits, each one with some probability of corresponding to the ground state. This distribution can be used in a fine-tuning loop, for example, reinforcement learning. In reinforcement learning, the framework takes a circuit from that distribution, runs it, and measures the energy. It feeds the results back into the transformer, which adjusts its distribution. Over time, the model learns to prioritize circuits that prepare increasingly accurate ground states.
Crucially, reinforcement learning allows the system to surpass its original training data instead of merely imitating it. The model is no longer acting as a compressed lookup table for ADAPT-VQE. It begins exploring novel circuit configurations that may outperform the teacher algorithm itself. This is one of the most important conceptual shifts in the project.
In this case, instead of running all the initial circuits on ĢƵ’s Helios, the reinforcement learning circuits were run using NVIDIA accelerated computing and the CUDA-Q platform, simulating a quantum processor.
Finally, once the transformers are optimized via reinforcement learning, the best resulting circuits are validated for accuracy and feasibility, by running them using InQuanto and on ĢƵ’s newest hardware, Helios. With InQuanto v5.2, users can now with both the Helios quantum computer and the Selene quantum emulator through Nexus.
This powerful combination of InQuanto and Nexus enabled the execution one of the largest AI-generated quantum chemistry circuits to date on a quantum computer; helping to turn the promise of quantum computing into a practical tool for pharmaceutical development.
Teaching a Transformer
Looking farther in the future, the researchers envision something much larger than a single molecular benchmark.
For bigger and more complex molecules, ADAPT-VQE won’t work in the first place as the initial training “oracle”. In addition, the molecular energy calculations used in the reinforcement learning grow too large for classical systems simulating quantum computers, so the quantum processor becomes essential.
Luckily, this is not a problem. The ultimate goal of the ADAPT-GQE framework is to develop a “curriculum” for the transformers. This means instead of re-training them for every new molecule, you instead keep what you already learned, and expand your knowledge from there.
By initially teaching it on molecules that are smaller, and that can be fully simulated, you ensure it learns on good data that can be double checked using known methods. From there, you can carefully build up the complexity to see how the transformer learns. Eventually, you hope to train it on molecules that can’t be simulated classically, using purely quantum data, all the while getting closer to the complexity levels you’re chasing.
This penultimate result is called a ‘foundation model’, which is a massive AI neural network trained on vast, broad datasets that can be adapted to a wide variety of downstream tasks. In this case, the team is building the very ‘foundations’ of a model that can solve the ‘electronic structure problem’, which is the core computational challenge lying at the heart of quantum (and classical) computational chemistry.
A New Interface Between AI and Quantum Computing
What makes this work particularly interesting is that it treats quantum circuit generation as a language modeling problem: circuits become sequences, transformers learn distributions over those sequences, and reinforcement learning optimizes them against physical reward functions.
The result is an AI system capable of proposing quantum circuits that were never explicitly programmed by humans.
That does not mean generative AI is replacing physics or chemistry. Instead, it is becoming a new interface layer for navigating unimaginably large search spaces that traditional optimization methods struggle to explore efficiently.
For quantum chemistry, that could become transformative.
If successful, frameworks like ADAPT-GQE may eventually allow researchers to synthesize useful quantum circuits for molecular systems too large for classical computation, accelerating everything from materials discovery to pharmaceutical design.
The broader implication is difficult to ignore: foundation models may eventually extend beyond language, images, and code — and into the fabric of physical reality itself.
