Increasingly a dominant challenge in our field has become coping with the enormous growth in complexity of data: its volume, its heterogeneity, its multiple messengers, its instrumental artifacts, and its astrophysical confounders.  Machine Learning and Artificial Intelligence (ML/AI) offer promise, but their key drawback has been the difficulty of interpreting the models.

 

In this talk I describe a new NASA Theoretical Computational Astrophysics Network (TCAN) program centered at UT Austin to overcome these challenges for JWST spectra of substellar objects.  I underscore the key enabling technology, “autodiff", as unlocking interpretability in ML/AI frameworks such as PyTorch and JAX.  This work includes an open-source, interpretable ML/AI ecosystem for astronomical spectroscopy.  I describe the “muler" and “gollum" Python APIs aimed at lowering the barrier to entry to e.g. IGRINS and HPF.  I illustrate this interpretability and scalability in our code blasé, that can tune >10^4 spectral lines simultaneously with autodiff in under 1 minute on NVIDIA GPUs.  I debut the “ynot" code, a prototype ''spectrograph digital twin'' that can enable next-generation Extreme Precision Radial Velocity (EPRV) experiments.  Finally, I propose dual-use technology development as a strategy for a new durable instrumentation program here at UT Austin.  I illustrate this strategy in Hydrogen leak detection and quantification satisfying both decarbonization and astronomical device development goals.  I show that open source, ML, and remote sensing tech development can derisk funding portfolios, including the new Simons Foundation Scientific Software Research Faculty Award, proposed here.