CDS&E: Accelerating Astrophysical Insight at Scale with Likelihood-Free Inference

在未来的调查中，以前所未有的深度和图像保真度，发现和分类的来源将是必要的，完全自动化，部分是通过使用机器学习。 该项目将开发新的算法和代码库，以加速时域天体物理学中基于模型的推理。 从计算和人力成本的角度来看，推理正在成为一个根本的瓶颈，参数估计任务可能是科学进步的主要障碍。 通常，领域专家需要监督拟合，以缩小搜索空间并加速推理。 这项工作将侧重于使用神经网络的似然自由推理（LFI）方法。 它将创建一个新的开源LFI Python库，并为引力微透镜和食双星执行特定领域的推理任务。 该软件将允许非专家使用LFI，计算代价较小，并在数据收集期间连续进行动态推理。 运行多学科研讨会并将LFI添加到本科和研究生课程中将改善研究基础设施和STEM教育。这种方法意味着计算前向模型的成本和时间在训练期间被覆盖，然后用于新数据的参数估计。这项工作将a）为不规则采样和噪声时间序列数据创建LFI架构，B）推进异常检测和后验校准方法，确保LFI参数估计是无偏的，以及c）开发快速模型选择方法。 飞行中的LFI可以为进一步的数据采集提供信息，这种良性的反馈循环有助于更有效地利用昂贵的后续资源。该奖项反映了NSF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。

In future surveys to unprecedented depth and image fidelity, the discovery and classification of sources will, by necessity, be fully automated, in part by using machine learning. This project will develop novel algorithms and codebases to accelerate model-based inference in time-domain astrophysics. Inference is becoming a fundamental bottleneck, from both a computation and human cost perspective, and the parameter estimation task can be the main impediment to scientific progress. Often, domain experts are required to supervise fitting in order to narrow search spaces and speed up inference. This work will focus on likelihood-free inference (LFI) approaches using neural networks. It will create a new open-source LFI Python library, and carry out domain-specific inference tasks for gravitational microlensing and for eclipsing binary stars. The software will let non-experts use LFI with smaller computational penalties, and enable on-the-fly inference continuously during data collection. Running a multidisciplinary workshop and adding LFI into undergraduate and graduate courses will improve both the research infrastructure and STEM education.This approach means that the cost and time to compute forward models are covered during training, and then leveraged for parameter estimation on new data. This effort will a) create LFI architectures for irregularly sampled and noisy time-series data, b) advance methods for anomaly detection and posterior calibration, ensuring LFI parameter estimates are unbiased, and c) develop approaches for fast model selection. On-the-fly LFI can inform further data acquisition, and this virtuous feedback loop helps with more efficient use of expensive follow-up resources.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.