Introduction
Solving the time-independent Schrödinger equation from first principles (ab initio) remains a fundamental challenge in quantum chemistry. Recently, Neural Network-based Variational Monte Carlo (NN-VMC) has demonstrated significant promise by using deep neural networks to represent complex electronic wavefunctions. However, a major computational bottleneck persists: the calculation of the Laplacian.
To determine the system's energy (the loss function), the model must compute the sum of second derivatives of the network's output with respect to electron coordinates. Standard deep learning frameworks typically compute the full Hessian matrix to extract the Laplacian trace, a process that is computationally redundant and prohibitively slow for large systems. For instance, simulating a benzene dimer can consume upwards of 10,000 GPU hours using standard methods.
The Forward Laplacian Framework
To address this inefficiency, researchers from Peking University and ByteDance Research introduced the Forward Laplacian framework. Unlike traditional "backward-then-forward" passes, this method utilizes a single forward propagation pass to compute the value, gradient, and Laplacian simultaneously for each layer.
By deriving a new chain rule for second-order derivatives, the framework avoids the $O(N^2)$ complexity associated with full Hessian matrices, where $N$ is the number of coordinates. This direct approach roughly halves the base computational cost before further optimizations are applied.
LapNet and Sparse Derivative Attention
Complementing the framework is LapNet, a specialized architecture designed to exploit physical properties of quantum systems. LapNet incorporates Sparse Derivative Attention (SDA) blocks. Recognizing that many physical interactions in quantum systems are local or decay exponentially, the derivatives of these functions are naturally sparse. LapNet preserves this sparsity during the forward pass, significantly reducing memory usage and increasing throughput.
Experimental Results and Performance
The framework was evaluated across diverse chemical benchmarks, including absolute energy for 16 molecules (up to 84 electrons), ionization potentials for transition metals, and reaction barrier heights from the BH76 dataset.
Key findings include:
- Significant Speedup: The combination of Forward Laplacian and LapNet achieved speedups of over 10x generally, reaching up to 20x for larger systems like long polyethylene chains $(C_2H_4)_n$.
- Linear Scaling: As shown in the scaling tests, the acceleration rate increases almost linearly with system size, suggesting high scalability for complex molecular simulations.
- High Accuracy: LapNet matched or exceeded the accuracy of state-of-the-art models like FermiNet and Psiformer, achieving "chemical accuracy" (errors within ~1.6 mHa) for reaction barriers and ionization potentials.
- Breaking the Size Barrier: The efficiency gains enabled the simulation of the uracil dimer (116 electrons) in approximately 1,800 GPU hours, a task estimated to take over 10,000 hours with previous methodologies.
Conclusion
The Forward Laplacian framework represents a transformative shift in NN-VMC, turning a primary computational bottleneck into a streamlined forward pass. By integrating physical insights like derivative sparsity into the network architecture, LapNet enables the study of large-scale biochemical systems with unprecedented efficiency. Future work may extend this framework to other domains governed by partial differential equations, such as fluid dynamics and heat diffusion.
Summary updated