More information
Description
The design of new materials by means of experiments is remarkably challenging due to the vast number of precursors and the difficulty of predicting the properties of novel materials prior their profound analysis. Theoretical modeling can assist experimental studies in efficiently devising novel compounds that feature desired properties. Unfortunately, conventional theoretical models are difficult primarily because the computational resources required grow exponentially with system size. Thus, highly accurate calculations at the quantum level are typically limited to small building blocks of larger materials. Novel computational methods can serve as a resort to break the unfavorable computational scaling of present-day models. One such innovative approach models many-electron systems as collections of electron pairs. Unfortunately, highly-optimized software packages that support such methods are currently unavailable. To allow for an efficient design of molecular compounds, the open-source software package PyBEST will be (i) extended to include an optimized tensor contraction engine and (ii) accelerated to support both CPUs and GPUs using modern approaches like Intel''s oneAPI and CuPy. The resulting optimized, open-source software suit will allow for an efficient design of organic solar cells (OSC) exploiting the robust, computationally cheap, reliable, and black-box-like methods shipped with PyBEST. These technical advantages compared to conventional codes and methods will facilitate theoretical modeling of molecules, which are out of reach of present-day quantum models. This project will shift the current paradigm in computational chemistry, large-scale modeling, and theoretical materials design towards novel and systematically improvable approaches implemented in modern codes that use progressive programming models. Finally, the advancement in the development of novel OSCs is particularly important for the renewable energy sector, not only in Poland but worldwide.
Summary of project results
The design of new materials by means of experiments is remarkably challenging due to the vast number of precursors and the difficulty of unambiguously predicting the properties of novel materials prior their profound analysis. Theoretical modelling can assist experimental studies in efficiently devising novel compounds that feature desired properties. Unfortunately, conventional state-of-the-art theoretical models are difficult, primarily because the computational resources required grow exponentially with system size. Thus, highly accurate quantum chemistry calculations are typically limited to small building blocks of larger materials. Novel electronic structure methods can serve as a resort to break the unfavorable computational scaling of present-day quantum chemistry. One such innovative approach models many-electron systems as collections of electron pairs or geminals. Unfortunately, highly-optimized quantum chemistry software packages that support geminal-based methods are currently unavailable. To allow for an efficient design of molecular compounds, our open-source software package PyBEST will be (i) extended to include an optimized tensor contraction engine and (ii) accelerated to support both CPUs and GPUs using modern approaches like Intel''s oneAPI and CuPy. The resulting optimized, open-source software suit will allow for an efficient design of organic solar cells exploiting the robust, computationally inexpensive, reliable, and black-box-like methods shipped with PyBEST. These technical advantages compared to conventional electronic structure codes and methods will facilitate theoretical modelling of molecules, which are out of reach of present-day quantum chemistry. Finally, this project will shift the current paradigm in computational chemistry, large-scale modelling, and theoretical materials design towards novel and systematically improvable approaches implemented in modern quantum chemistry codes that use progressive programming models.
The advancement of light-harvesting materials and highly-efficient electroluminescent devices is of growing importance in both academia and industry. However, in order to compete with other (photovoltaic) technologies, further improvements are still desirable. This can be achieved by means of more reliable quantum chemical predictions of electronic structures and properties of modern materials. Yet, the size of common OSC components prohibits the use of standard wave-function methods, whereas DFT might fail due to their MR nature. Because of their efficiency and good performance, pCCD-based methods can be employed to, for instance, effectively screen and group various OSC motifs, providing a one-of-a-kind data set library for the future development of electronic devices and predict (not simply reproduce) experimental data that can be exploited in molecular design. Thus, this project may shift the current paradigm in computational chemistry, large-scale modeling, and theoretical materials design towards novel and systematically improvable approaches (beyond DFT) implemented in modern quantum chemistry codes that use progressive programming models. Furthermore, the outcome of the project can provide a reliable database of OSCs for machine learning. However, to make pCCD-based methods applicable to model OSC components, additional improvements are required: (1) the computer implementations have to allow us to study large systems efficiently, (2) the mathematical models have to be reliable and the corresponding optimization algorithms robust, and (3) the underlying software has to be user-friendly. This project aims at achieving these goals by (1) offloading the bottleneck operations to the GPU. The final GPU-accelerated linear algebra library allows for a speed-up of a factor of 3 to 4 compared to the CPU-only implementation. (2) The pCCD models have been extended to significantly improve the accuracy and reduce errors with respect to experimental results. (3) A GUI to the PyBEST software package allows users to construct input files for large-scale simulations using a cross-platform application. This project represents a step towards breaking the computational paradigm of quantum chemical modeling of organic electronics by bridging novel quantum chemistry approaches with modern programming models like GPU acceleration and user-friendly cross-platform graphical interfaces.
We demonstrated that Pythonic coupled cluster implementations can be efficiently and straightforwardly offloaded to the GPU using the CuPy library. A CuPy version of selected NumPy routines may perform similarly to pure CUDA implementations if the problem under study completely fits on the video RAM. Otherwise, a performance decrease is to be expected as passing array slices in Python to the GPU requires looping. Nonetheless, these drawbacks can be minimized if the video RAM is increased or NVLink switch technology is used for multi-GPU utilization. Our findings motivate further development in GPU-accelerated Python-based computing exploiting other libraries, like cutensor, PyTorch, or opt_einsum’s GPU backend. The flexible and improved structure of the linear algebra library in PyBEST facilitates an extension of our GPU support to other libraries than CuPy. This will make PyBEST competitive and resilient for new features and future dev options. GPU-accelerated computing in PyBEST has been provided for all bottle-neck operations in most coupled-cluster-type methods, including those based on the pair coupled cluster doubles (pCCD) model. To improve the performance of pCCD-based approaches in terms of predicting reliable molecular properties, we had to extend the existing models and improve the underlying solvers to ensure robust and smooth convergence. Our numerical studies (using the optimized solvers and tensor contractions) indicate that pCCD-based models are a powerful alternative to model organic electronics as they provide many tools to facilitate their description and interpretation, like orbital energies, localized orbitals (to clearly distinguish charge-transfer states from locally-excited or Rydberg states), quantum embedding, and quantum entanglement and correlation. Unfortunately, to outperform present state-of-the-art coupled cluster approaches (of similar computational complexity), additional advancements are required. We demonstrated that higher excitations (triples, etc.) are important if we wish to approach chemical accuracy in excited states or their molecular properties. Finally, our cross-platform graphical user interface allows users to easily set up a pCCD or post-pCCD calculation without any substantial knowledge of Python. The developed PyBEST-GUI is also beneficial for training young researchers or students (across disciplines) in unconventional electronic structure methods. All methodological and software developments mentioned above are available free of charge on various repositories (PyPI, zenodo), which makes them findable, accessible, interoperable, and reproducible (FAIR).