AI Designs Quantum Algorithms for Molecules LiH, H2O

Fabian Finger and colleagues at Quantinuum, in a collaboration between Quantinuum and Hiverge, have created Hive, an AI platform that automatically discovers new algorithms for solving complex computational problems. The platform determines the ground state energy of molecules, specifically LiH, H2O, and F2, and provides a sharp reduction in the quantum resources needed compared with existing approaches. Benchmarking these AI-discovered circuits on the Quantinuum System Model H2 establishes minimum hardware requirements for achieving chemical accuracy and suggests this method could extend beyond chemistry to other areas of quantum computation.

AI-driven algorithm discovery substantially reduces qubit requirements for molecular energy

Up to 70% fewer two-qubit gates were needed for LiH, H2O, and F2 molecules when determining ground state energies, compared to state-of-the-art variational quantum eigensolver (VQE) algorithms. This reduction unlocks calculations on near-term quantum computers previously beyond their capabilities, achieving chemical accuracy and results comparable to classical methods with fewer qubits and circuit evaluations. The ground state problem, central to quantum chemistry, involves finding the lowest energy state of a molecule, a computationally demanding task that scales exponentially with system size using classical computers. VQE algorithms represent a leading approach for tackling this problem on near-term quantum devices, but their performance is heavily reliant on the design of the quantum circuit, or ansatz, used to approximate the molecular wavefunction. Hive, an AI platform, autonomously discovered these algorithms, opening avenues for wider application of quantum computation and bypassing the need for expert intuition in quantum circuit design. Traditional ansatzes, such as the Unitary Coupled Cluster (UCC) ansatz, while theoretically powerful, often require a significant number of qubits and circuit depth, exceeding the capabilities of current hardware. The reduction in gate count achieved by Hive is particularly significant as two-qubit gates are generally the most error-prone operations on near-term quantum computers.

Hive iteratively refined algorithms by employing program synthesis and a distributed evolutionary process, presenting a novel approach to quantum algorithm discovery applicable beyond chemistry and potentially extending to fault-tolerant quantum computers. Program synthesis involves automatically generating computer programs from a high-level specification, in this case, the goal of minimising the energy of a molecule. The evolutionary process mimics natural selection, where algorithms are evaluated based on their performance, and the best-performing algorithms are ‘bred’ together to create new generations of algorithms. This process is highly distributed, utilising a network of computational resources to explore a vast search space of possible algorithms. Calculations on the LiH molecule, for example, required fewer than 50 circuit evaluations, establishing minimum system requirements for achieving chemical precision on the Quantinuum System Model H2 quantum computer. Chemical accuracy is typically defined as achieving an energy error of less than 1 kcal/mol, equivalent to approximately 4.184 kJ/mol. The platform also identified key functions responsible for efficiency gains, revealing insights into how quantum circuits can be optimised for performance. These identified functions suggest that certain circuit structures and gate arrangements are particularly effective at representing the molecular wavefunction with minimal resources.

The AI platform successfully tackled the ground state problem for H2O and F2 molecules, demonstrating flexibility beyond simple diatomic systems. This automated approach contrasts with traditional methods reliant on expert intuition, potentially accelerating quantum algorithm development, although current results focus on relatively small molecules. The ability to generalise to more complex molecules is crucial for demonstrating the practical utility of this approach. Water and fluorine, while still relatively small, present a greater challenge than lithium hydride due to their more complex electronic structures and increased number of atoms. Further investigation is now prompted into how the identified efficiency-driving functions can inform both AI-driven and human-guided optimisation strategies. Understanding the underlying principles behind the AI’s discoveries could lead to the development of new, more efficient quantum algorithms designed by human experts, creating a synergistic relationship between AI and human ingenuity.

Scaling AI-designed quantum algorithms to complex molecular systems

Automating the design of quantum algorithms promises to unlock solutions to previously intractable problems in fields such as materials science and drug discovery. The computational cost of simulating molecular properties limits the size and complexity of systems that can be studied using classical methods. Quantum computers offer the potential to overcome these limitations, but require efficient algorithms to harness their power. This achievement for lithium hydride, water, and fluorine signifies a potential shift towards AI-augmented quantum computation, lessening the need for expert intuition in algorithm development. The Hive platform, utilising program synthesis and large language models, independently discovered algorithms for determining molecular ground state energies, offering a new approach to solving complex computational problems and bypassing limitations inherent in manual circuit creation. Large language models are employed to guide the search process, leveraging their ability to understand and generate complex patterns in data. The team acknowledges that demonstrating consistent performance improvements on increasingly challenging molecular structures remains essential for validating the long-term potential of this approach. Future work will focus on scaling these algorithms to larger molecules, such as peptides and proteins, which are relevant to drug discovery and materials science.

The current work represents a significant step towards automating the quantum algorithm design process, but several challenges remain. One key challenge is the exploration of the vast algorithm space. The number of possible quantum circuits grows exponentially with the number of qubits, making it computationally intractable to search all possibilities. Hive addresses this challenge by employing a distributed evolutionary process, but further improvements in search efficiency are needed. Another challenge is the development of robust evaluation metrics. Accurately assessing the performance of a quantum algorithm requires careful consideration of factors such as circuit depth, qubit count, and error rates. The development of more sophisticated evaluation metrics will be crucial for guiding the AI towards the discovery of truly optimal algorithms. Finally, the integration of domain-specific knowledge into the AI platform could further enhance its performance. For example, incorporating knowledge of chemical bonding and molecular structure could help the AI to design more efficient algorithms for specific types of molecules. The ultimate goal is to create an AI platform that can autonomously design quantum algorithms for a wide range of scientific and engineering problems, accelerating the development of quantum technologies and unlocking their full potential.

The research successfully used artificial intelligence, specifically the Hive platform, to discover new quantum algorithms for calculating the ground state energy of molecules like lithium hydride, water, and fluorine. This matters because designing these algorithms manually is difficult, and AI-driven discovery could lead to more efficient quantum computations with fewer quantum resources. The discovered circuits were even tested on Quantinuum’s System Model H2 computer, identifying minimum hardware requirements for accurate results. This approach potentially extends beyond chemistry and could facilitate the design of algorithms for more powerful, fault-tolerant quantum computers, ultimately aiding advancements in fields like drug discovery and materials science.