2026-07-15 –, University Hall
Many scientific workflows rely on derivatives of complex models: Fisher forecasts, sensitivity analysis, gradient-based inference, and emulator construction. In practice, these derivatives are often difficult to compute reliably and integrate into end-to-end inference pipelines.
DerivKit is an open-source Python toolkit that provides a unified framework for derivative-based scientific inference. It supports multiple derivative backends and connects model evaluation directly to downstream inference tools, including Fisher analyses and higher-order likelihood approximations. The framework also provides diagnostics and visualization tools for exploring parameter sensitivities and degeneracies.
Originally developed for cosmological forecasting pipelines, DerivKit is designed to be domain-agnostic and easily integrated into scientific Python workflows.
Many scientific workflows rely on derivatives of complex computational models. Derivatives are central to Fisher forecasting, sensitivity analysis, gradient-based inference, emulator construction, and uncertainty propagation. In practice, however, derivative calculations are often implemented in ad-hoc ways within individual projects. This makes them difficult to reproduce, hard to diagnose when they fail, and challenging to integrate with downstream inference tools.
DerivKit is an open-source Python toolkit designed to provide a structured framework for derivative-based scientific inference. The goal of the project is to connect model evaluation, derivative computation, and inference tools into a coherent workflow that is easy to use and inspect. Rather than focusing on a single derivative technique, DerivKit provides a unified interface for multiple derivative backends and supports flexible strategies for computing derivatives of arbitrary scientific models.
The framework allows users to wrap an existing model function and automatically construct derivative operators with respect to model parameters. These derivatives can then be used directly in inference pipelines, including Fisher matrix forecasts and higher-order likelihood approximations (DALI). In particular, DerivKit provides implementations of higher-order likelihood expansions that extend beyond the Gaussian Fisher approximation, enabling users to explore parameter degeneracies and non-Gaussian structure in likelihood surfaces.
An important design goal of DerivKit is to make derivative-based inference transparent and diagnostic-friendly. The toolkit includes utilities for evaluating derivative stability, exploring parameter sensitivities, and visualizing degeneracies in model parameter spaces. These diagnostics help users identify when derivatives are unreliable or when parameter combinations produce nearly degenerate model responses. DerivKit also supports a direct model-to-plot workflow that allows users to move seamlessly from derivative computation to visual analysis of inference results.
Although DerivKit was originally developed for cosmological forecasting pipelines used in large astrophysical collaborations, the design of the framework is intentionally domain-agnostic. Many areas of scientific computing face similar challenges when working with derivatives of expensive or complex models. These include climate modeling, epidemiological simulations, materials science, and simulation-based inference workflows. By separating derivative infrastructure from domain-specific modeling code, DerivKit aims to provide a reusable tool that can integrate naturally into a wide range of scientific Python environments.
This talk will introduce the design principles behind DerivKit and demonstrate how derivative infrastructure can be organized to support robust scientific inference workflows. We will discuss common pitfalls in numerical derivative calculations, present the architecture of the DerivKit framework, and show examples of derivative-based inference applied to realistic models.
Attendees will learn how to structure derivative computations in a reproducible way, how to diagnose instability and parameter degeneracies, and how derivative-based methods such as Fisher analyses and higher-order likelihood approximations can be incorporated into scientific Python pipelines.
Nikolina “Niko” Šarčević is a cosmologist at Duke University working on cosmological inference and large-scale structure. She is a member of the LSST Dark Energy Science Collaboration (DESC) and the NASA Roman Space Telescope science collaborations. Her research focuses on statistical methods, astrophysical systematics, and scientific software for cosmology. Previously, she worked on dark matter searches as part of the XENON experiment.
