Yang Ming Chiao Tung University Team Introduces Hi-Nqs Algorithm for Correlated Systems

Scientists are continually striving to accurately solve the electronic Schrödinger equation for strongly correlated systems, a central challenge in quantum chemistry where the computational cost escalates exponentially with system size. Jen-Yu Chang and colleagues at Yang Ming Chiao Tung University, in collaboration with the National Centre for High-performance Co, Chung Yuan Christian University, and the National Institutes of Applied Research, have developed a new algorithm, the Handover Iterative Neural Quantum State (HI-NQS), which combines the strengths of neural networks and traditional quantum chemistry methods. This purely classical approach utilises a Transformer neural quantum state within an iterative framework to efficiently construct compact and chemically accurate wavefunction representations, demonstrating sharply improved scaling compared to conventional methods for all but the smallest systems tested.

Reduced scaling and enhanced accuracy in quantum chemical simulations using a novel neural network

A new Handover Iterative Neural Quantum State (HI-NQS) algorithm achieves a log-linear determinant count scaling exponent of 0.089 ±0.011. This represents a significant advancement over existing methods, achieving roughly half the scaling of conventional CIPSI-based Selected Configuration Interaction (SCI) methods and approximately one-third that of Complete Active Space Configuration Interaction (CASCI). The importance of this reduction in computational cost lies in its ability to enable simulations of systems with up to 40 qubits, equivalent to approximately 6x 10 9 determinants, which were previously intractable for accurate modelling. Strongly correlated systems, where electron interactions are particularly significant, require exponentially growing computational resources with traditional methods. The HI-NQS algorithm achieved chemical accuracy, defined as energy errors below 1.6x 10 -3 Ha, across a diverse range of molecules and nitrogen active-space series, consistently surpassing the performance of existing techniques. For instance, in the case of the H 2 S molecule, HI-NQS required smaller basis sets of determinants, reducing the count by a factor of 0.066, representing a fifteen-fold decrease in computational demand. This reduction allows for more efficient exploration of potential energy surfaces and more accurate prediction of molecular properties. Across nitrogen active-space calculations, the algorithm’s determinant count scaled with a log-linear exponent of 0.089 ±0.011, sharply outperforming both CIPSI-SCI (0.180 ±0.012) and CASCI (0.262 ±0.006). These scaling exponents are crucial as they dictate how computational cost increases with system size; a lower exponent signifies a more efficient algorithm. Current implementation relies on GPU hardware for accelerated computation and does not yet address scaling to even larger systems or integration with quantum computing resources, representing potential avenues for future research.

Handover Iterative Neural Quantum States for Molecular Energy Calculation

Computational limitations in quantum chemistry, particularly when dealing with strongly correlated systems, were overcome by developing a technique that iteratively refines understanding of a molecule’s electronic structure. The core of this approach is a ‘Transformer neural quantum state’, an artificial intelligence model inspired by the principles of natural language processing. Transformers excel at identifying patterns and relationships within complex data, and in this context, they learn to represent the quantum state of a molecule. This is analogous to how a computer learns to predict the next word in a sentence, but instead of words, the model deals with the complex mathematical functions that describe electron behaviour. Rather than directly solving the electronic Schrödinger equation, the fundamental equation in quantum mechanics describing how a system changes over time, the model proposes building blocks for the solution. These building blocks, known as determinants, represent possible configurations of electrons within the molecule. Researchers at Chung Yuan Christian University developed the HI-NQS algorithm to calculate molecular energy without using quantum computers, employing this approach to represent a molecule’s electronic structure and iteratively improve its understanding. The iterative process involves repeatedly refining the neural network’s representation of the wavefunction, guided by the principles of quantum mechanics, until a sufficiently accurate solution is obtained. This allows the algorithm to focus on the most important determinants, significantly reducing the computational burden.

Addressing wavefunction collapse in a novel classical molecular modelling algorithm

An algorithm, HI-NQS, has been unveiled by researchers at Yang Ming Chiao Tung University and collaborating institutions, offering a compelling alternative for modelling complex molecular systems. The approach improves efficiency in constructing wavefunctions, the mathematical description of a molecule’s electronic structure, by intelligently selecting the most important configurations. However, its reliance on an autoregressive sampler introduces a potential vulnerability to ‘mode collapse’. This phenomenon, common in generative models, occurs when the algorithm fixates on a limited set of solutions, failing to adequately represent the full complexity of the system. This could hinder accurate representation of systems exhibiting intricate behaviours, particularly those involving broken symmetry (where molecules lack certain symmetries) or dissociation (where molecules break apart). Combining a Transformer neural quantum state with an iterative refinement process efficiently builds solutions to the complex electronic Schrödinger equation, achieving chemical accuracy, a standard for reliable molecular simulations, while substantially reducing computational effort compared to existing techniques. This efficiency stems from focusing on the most chemically relevant parts of a molecule’s electronic structure, offering a purely classical route to accurately modelling strongly correlated systems and sidestepping the need for quantum computers. The ability to accurately model strongly correlated systems has implications for a wide range of fields, including materials science, drug discovery, and catalysis, where understanding electron interactions is crucial for predicting and controlling chemical behaviour. Further research will focus on mitigating the risk of mode collapse and extending the algorithm’s scalability to even larger and more complex systems.

The researchers developed a new algorithm, HI-NQS, which accurately solves the electronic Schrödinger equation for complex molecular systems. It achieves this by efficiently constructing wavefunctions, focusing on the most chemically important configurations and reducing computational demands compared to conventional methods. This approach offers a purely classical means of modelling strongly correlated systems, demonstrating chemical accuracy across a series of small molecules and nitrogen-based active spaces. The authors intend to address potential limitations related to ‘mode collapse’ and improve the algorithm’s ability to handle even larger molecular systems.