Chemical Simulations Boosted with New Technique Achieving 94% Efficiency

Scientists are increasingly reliant on high-performance computing to tackle complex chemical problems involving large systems or demanding high accuracy. Qingpeng Wang, Ning Zhang, and Wenjian Liu, all from the Qingdao Institute for Theoretical and Computational Sciences and Shandong University, have now developed a unified Message Passing Interface (MPI) parallelisation strategy for wave function methods, demonstrated using the iCIPT2 method. This work significantly advances computational chemistry by abstracting each step of a method as a dynamically-scheduled loop, enabling a single MPI template applicable to diverse calculations and achieving exceptional parallel efficiencies of up to 94% on 1024 cores. The resulting computational power facilitates benchmark calculations for challenging systems like cyclobutadiene, benzene, and ozone, and reveals a predictable scaling behaviour of iCIPT2 with respect to configuration state functions.

Unified MPI parallelisation streamlines high-accuracy quantum chemical calculations

Scientists have achieved a significant advance in computational chemistry through a unified approach to parallelizing complex wave function methods. This work introduces a novel MPI (message passing interface) parallelization strategy applicable across diverse quantum chemical calculations, demonstrated effectively using the iCIPT2 method.
The research addresses a longstanding challenge in handling large chemical systems with high accuracy, a feat previously limited by computational cost. By abstracting each computational step as a dynamically-scheduled loop and employing a ‘ghost process’ for data distribution, researchers have created a versatile and efficient parallelization template.

This algorithmic abstraction allows a single MPI template to be used across various steps within different methods, streamlining the parallelization process. Benchmarking the iCIPT2 method on 16 nodes, comprising 1024 cores, revealed parallel efficiencies of 94% for perturbation calculations and 89% for complete calculations.
These results signify a substantial improvement in computational performance, enabling calculations previously inaccessible due to their complexity. Further algorithmic enhancements, including an improved matrix-vector product and a semi-stochastic estimator, facilitate calculations on large active spaces.

Consequently, the team obtained benchmark data for the automerization of cyclobutadiene, determined the ground state energy of benzene, and mapped the potential energy profile of ozone. The study also establishes a power law relationship between the error of iCIPT2 and the number of configuration state functions, providing valuable insight into the method’s accuracy and scalability. This unified MPI parallelization, coupled with algorithmic refinements, represents a crucial step towards tackling increasingly complex chemical problems with unprecedented precision.

Implementation of a dynamically-scheduled MPI parallelisation for wave function calculations

A unified MPI parallelization strategy underpinned this work, abstracting each computational step of wave function methods as a dynamically-scheduled loop utilising ghost processes. This approach facilitated a global reduction of local results from each node, enabling efficient distribution of computational load.

The algorithmic abstraction proved crucial, allowing a single MPI template to be applied across diverse steps within different methods, streamlining the parallelisation process. iCIPT2 served as a demonstrative case, achieving parallel efficiencies of 94% and 89% on 16 nodes, comprising a total of 1024 cores, for the perturbation and complete calculations respectively. An improved algorithm for the matrix-vector product within matrix diagonalisation was implemented, accelerating this critical step.

Furthermore, an orbital-configuration-based semi-stochastic estimator was developed for the perturbation correction, enabling calculations with larger active spaces. These methodological advancements facilitated benchmark calculations for the automerization of cyclobutadiene, determination of the ground state energy of benzene, and mapping the potential energy profile of ozone.

The study also demonstrated that the error associated with iCIPT2 scales according to a power law relative to the number of configuration state functions, providing insight into the method’s accuracy and limitations. The research leveraged the MetaWave platform, building upon its unified implementation of both non-relativistic and relativistic wave function methods, as previously detailed in J.

Phys. Chem. A0.2025, 129, 5170. This integration, combined with the tabulated unitary group approach and orbital configuration pair-based classification, allowed for rapid evaluation and reuse of key coupling coefficients between configuration state functions.

High-performance parallelisation of iCIPT2 calculations for accurate molecular energetics

Parallel efficiencies of 94% and 89% were achieved on 16 nodes, comprising 1024 cores, for the perturbation and complete calculations, respectively, utilising the iCIPT2 method. This unified MPI parallelization was accomplished by abstracting each computational step as a dynamically-scheduled loop with ghost processes, followed by global reduction of local results from each node.

The algorithmic abstraction facilitated the implementation of a single MPI template across various steps within different methods, streamlining the parallelization process. An improved algorithm for the matrix-vector product during matrix diagonalization, combined with an orbital-configuration-based semi-stochastic estimator for perturbation correction, enabled calculations within large active spaces.

Benchmarks were obtained for the automerization of cyclobutadiene, the ground state energy of benzene, and the potential energy profile of ozone, demonstrating the method’s applicability to complex chemical systems. These calculations provide detailed energetic data for these processes, furthering understanding of their underlying mechanisms.

The error associated with iCIPT2 calculations was shown to follow a power law relationship with respect to the number of configuration state functions. This scaling law provides valuable insight into the convergence behaviour of the method and allows for prediction of accuracy based on the size of the configuration space. The research establishes a foundation for efficient and scalable electronic structure calculations, paving the way for investigations of increasingly complex molecular systems and chemical phenomena.

High performance iCIPT2 calculations and validation against benchmark systems

Scientists have developed a unified implementation of wave function methods using message passing interface (MPI) parallelization, enabling calculations on large systems with improved accuracy. This was achieved by abstracting each computational step as a dynamically-scheduled loop, facilitated by a ghost process for efficient load balancing and a global reduction of local results.

The methodology was demonstrated using the iCIPT2 method, achieving parallel efficiencies of 94% and 89% on 16 nodes with 1024 cores for perturbation and complete calculations respectively. Further algorithmic improvements, including an optimised matrix-vector product and a semi-stochastic estimator, allowed for calculations previously inaccessible, such as benchmarks for the automerization of cyclobutadiene, the ground state energy of benzene, and the potential energy profile of ozone.

Analysis revealed that the error within iCIPT2 calculations scales predictably with the number of configuration state functions, following a power law relationship. The unified MPI template developed is readily adaptable to relativistic wave function methods, although current memory limitations related to data replication across nodes require further attention.

The authors acknowledge a memory bottleneck stemming from replicating configuration interaction vectors on each node and are actively working to address this by implementing a cross-node distribution of data. Future research will focus on removing this limitation to further enhance the scalability and efficiency of relativistic calculations. This work, supported by the National Natural Science Foundation of China, establishes a robust and versatile framework for high-performance quantum chemical computations.