CDS&E: D3SC: Accelerating Density Functional Theory Based Simulations and Materials Design with Machine Learning

乔治亚理工学院的Rampi Ramprasad和Le Song获得了化学系化学理论、模型和计算方法项目的奖励，以加速分子和材料的计算机模拟。该奖项由材料研究部凝聚态物质和材料理论项目共同资助。基于量子力学的模拟是准确的、通用的，对分子和材料的发现至关重要。不幸的是，对于许多现实世界的系统来说，它们也非常慢。因此，科学家常常局限于研究小尺度的理想系统。Ramprasad和Song正在利用人工智能方法开发一种新的、变革性的能力，这种能力可以观察、提炼和学习量子力学方程的行为以及化学（非）相似性的概念。他们的新“机器学习”能力能够模拟量子力学模拟，但在保持精度的同时，其速度超过了8个数量级。这种能力在质量上逐步提高，因为它定期（或按需）暴露在知识贫乏的区域的新化学和分子构型的新数据中。换句话说，仿真能力可以随着经验系统地、持续地提高。目前一些传统仿真方法无法解决的问题可能会得到解决。通过他们的研究、教学和传播努力，Ramprasad和Song正在培养下一代学生，他们将需要传统STEM领域的技能以及高级数据分析。该项目与美国国家科学基金会的“利用数据革命大构想”（Harnessing The Data Revolution Big Idea）非常契合。这一努力也与材料基因组计划（MGI）保持一致，该计划推动了加快新化学品和材料发现周期的努力，以满足各种关键技术需求。Ramprasad和Song在他们的工作中所针对的量子力学模拟的特定风格涉及解决密度泛函理论（DFT）的Kohn-Sham （KS）方程。尽管DFT具有通用性，但由于KS方程的计算瓶颈，常规DFT计算通常仅限于几百个原子。Ramprasad和Song正在开发的基于机器学习的DFT仿真方案利用过去的示例参考DFT计算有效地吸收了KS方程的函数。一旦经过训练，该方案就可以在未来的计算中完全绕过KS方程，直接、快速、准确地预测分子或材料的电子结构，只需要它的原子构型。利用新颖的旋转不变图神经网络表示将网格点周围的原子环境映射到该网格点的电子密度和局域电子密度。这种映射是使用在数百万个网格点的初始参考DFT结果上训练的深度神经网络来学习的。随后出现的新数据将通过多任务学习和迁移学习被神经网络吸收。该范例允许对KS DFT进行高保真仿真，但比直接解决方案快几个数量级；实际上，机器学习预测方案严格地与系统大小成线性比例。这种能力的目标是涉及碳、氢、氧和氮的系统，包括技术上重要的聚合物和分子的大型化学空间。其愿景是能够有效地设计具有目标电子结构的分子和聚合物。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Rampi Ramprasad and Le Song of the Georgia Institute of Technology are supported by an award from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry to accelerate computer simulations of molecules and materials. The award is co-funded by the Condensed Matter and Materials Theory program in the Division of Materials Research. Simulations based on Quantum Mechanics are accurate, versatile and vital to molecular and materials discovery. Unfortunately, for many real-world systems, they are also painstakingly slow. Thus, scientists are often limited to studying idealized systems at small scales. Ramprasad and Song are utilizing Artificial Intelligence approaches to develop a new, transformative capability that observes, distills and learns the behavior of the Quantum Mechanical equations and notions of chemical (dis)similarity. Their new "Machine Learning" capability is able to emulate Quantum Mechanical simulations, but at unprecedented speed-ups of over eight orders of magnitude while still preserving accuracy. This capability progressively improves in quality as it is periodically (or on-demand) exposed to fresh data on new chemistries and molecular configurations in regions of poor knowledge. In other words, the simulation capability can systematically and continuously improve with experience. Several problems currently beyond the reaches of traditional simulation methods may become solvable. Through their research, pedagogical and dissemination efforts, Ramprasad and Song are training the next generation of students who will require skills in traditional STEM domains as well as in advanced data analytics. The project is well-aligned with NSF's Harnessing the Data Revolution Big Idea. This effort is also aligned with the Materials Genome Initiative (MGI) which has propelled efforts to hasten the discovery cycle of new chemicals and materials to satisfy a variety of critical technological needs. The specific flavor of Quantum Mechanical simulations that Ramprasad and Song are targeting in their work involves solving the Kohn-Sham (KS) equation of density functional theory (DFT). Despite its versatility, routine DFT calculations are usually limited to a few hundred atoms due to the computational bottleneck posed by the KS equation. The Machine Learning based DFT-emulation scheme that Ramprasad and Song are developing assimilates the function of the KS equation efficiently using past example reference DFT calculations. Once trained, the scheme by-passes the KS equation entirely for future calculations, to directly, rapidly and accurately predict the electronic structure of a molecule or material, given just its atomic configuration. Novel rotationally invariant graph neural network representations are being utilized to map the atomic environment around a grid point to the electron density and local electronic density of states at that grid point. This mapping is learned using a deep neural network trained on initial reference DFT results at millions of grid points. Subsequently emerging new data will be absorbed into the neural networks via multi-task learning and transfer learning. The paradigm allows for the high-fidelity emulation of KS DFT, but orders of magnitude faster than the direct solution; indeed, the machine learning prediction scheme is strictly linear-scaling with system size. This capability is being aimed at systems involving C, H, O and N, which encompasses a large chemical space of technologically important polymers and molecules. The vision is to enable a capability to efficiently design molecules and polymers with a target electronic structure.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.