Goldilocks convergence tools and best practices for numerical approximations in Density Functional Theory calculations

在材料和分子科学领域，基于密度泛函理论(DFT)的建模和模拟是绿色计算、环境修复、能源生产、转换和储存等环境可持续性功能材料研发的关键。基于DFT的研究目前在全球超级计算机上消耗了大量资源。在英国，DFT计算使用了超过45%的ARCHER2，这是英国国家超级计算服务的第一级。DFT还在使用第二级系统和较低级别的机构计算机方面发挥了重要作用。随着越来越强大的计算机的出现，基于DFT的研究对环境的影响正在迅速增加。最重要的是提高这项研究的效率，并开发确保以负责任的方式分配和使用能源密集型计算资源的方法。拟议的工作将为研究人员和计算资源分配链提供实现这些目标的实用工具和基于证据的最佳实践。DFT计算包含需要根据每项研究所需的精度收敛的数值近似。如果不为经验不足的用户提供更多支持，信息系统就有过度收敛的风险，导致不必要的更昂贵的计算，或者收敛不足，导致完全无用的计算，这是对计算资源和电力的浪费。对欠收敛或过收敛的DFT计算比例的保守估计在10%的范围内。考虑到在这项研究中投入的计算资源比例很大，即使相对较小的效率提高也将导致浪费的计算资源大幅减少，并显著改善研究基础设施的环境可持续性。该项目将产生一个工具和基于证据的最佳实践，以提供自动的、专家指导的‘金发女孩’选择这些收敛参数。这将通过训练机器学习(ML)模型来实现，以预测DFT数值近似的收敛参数，以达到常见类型的科学调查所需的精度。鉴于所有代码中都存在需要收敛的数值近似，此工具将适用于英国常用的所有DFT代码。该项目的主要贡献将是显著提高UKRI和EPSRC硬件和软件基础设施的负责任使用的效率和保证水平。通过比较采用此工具前后运行的典型作业的计算资源使用情况，可以对获得的效率进行基线量化和外推。这一分析的结果将在全球传播，在国际计算设施中产生最佳做法，从而在世界范围内推广计算基础设施在环境可持续性方面的成果。该项目将是朝着基于ML的DFT计算投入的自动生成以及计算资源和碳足迹的自动先验计算器迈出的重要一步。这种自动化将有助于在世界某些地区使用这种研究方法的民主化，在这些地区，数字研究基础设施可能比实验设施更容易获得。

Within the field of materials and molecular science, modelling and simulation based on Density Functional Theory (DFT) is key in the R&D of functional materials for environmental sustainability, such as green computing, environment remediation, and energy production, conversion and storage. DFT-based research currently consumes a considerable amount of resources on supercomputers globally. In the UK, DFT calculations use over 45% of ARCHER2, the Tier1 UK National Supercomputing service. DFT also features heavily in the usage of Tier2 systems and lower-tier institutional computers. As ever more powerful computers become available, the environmental impact of DFT-based research is increasing rapidly. It is paramount to improve the efficiency of this research and develop means of assuring that energy-intensive compute resources are distributed and used responsibly. The proposed work will provide practical tools and evidence-based best practices towards these aims for researchers and the compute-resources distribution chain.DFT calculations contain numerical approximations that need to be converged according to the accuracy required for each study. Without more support for inexperienced users, the risk of is of over-convergence, leading to unnecessarily more costly calculations, or under-convergence, leading to entirely useless calculations, which are a waste of compute resource and electricity. A conservative estimate of the proportion of under- or over-converged DFT calculations is in the 10% range. Given the large proportion of compute resource invested in this research, even a relatively small increase in efficiency will result in a large reduction of wasted compute resource, and significant improvements in the environmental sustainability of research infrastructure.This project will result in a tool and evidence-based best practices to provide automatic, expert guiding in the 'Goldilocks' choice of these convergence parameters. This will be achieved by training machine learning (ML) models to predict the convergence parameters for DFT numerical approximations for the required accuracy in common types of scientific investigations. Given that numerical approximations requiring convergence are present in all codes, this tool will be applicable across all DFT codes in common use in the UK. The primary contribution of this project will be to increase considerably the efficiency and assurance levels of responsible use of UKRI and EPSRC hardware and software infrastructure, now and in the future.Comparison of the compute resources usage for typical jobs run before and after the adoption of this tool, will enable baseline quantification and extrapolation of the efficiency gained. Outcomes of this analysis will be disseminated globally, leading to best practices across international compute Facilities, so as to extend world-wide the gains in environmental sustainability of compute infrastructure.This project will be a significant step towards ML-based automatic generation of inputs for DFT calculations, as well as an automatic a priori calculator of compute resources and carbon footprint. This automation will contribute to democratisation in the use of this research method in parts of the world where digital research infrastructure may be more accessible than experimental facilities.