Compositional simulation-based inference for time series

Manuel Glöckler, Shoji Toyota, Kenji Fukumizu, Jakob H. Macke

Abstract

Simulation-based inference (SBI) is a powerful tool for parameter estimation in models with intractable likelihoods. However, applying SBI to time series data remains challenging due to the high dimensionality of sequences and the computational cost of full-trajectory simulations. We propose Factorized Neural Score Estimation (FNSE), a compositional approach that exploits the Markovian structure of many scientific simulators. By training on single-state transitions and aggregating local scores, FNSE achieves superior simulation efficiency and scales to high-dimensional problems, such as fluid dynamics simulations with over a million data dimensions.

Introduction

In many scientific domains, simulators are used to model complex phenomena, but calculating the likelihood of observed data is often impossible. Simulation-Based Inference (SBI) addresses this by training neural networks to approximate the posterior distribution of parameters given observations. For time series, standard SBI methods typically process entire sequences, which is computationally expensive and difficult to scale. Crucially, these methods often ignore the Markovian property—where the next state depends only on the current state and parameters—which is a fundamental characteristic of many physical and biological models.

Methodology

The proposed Factorized Neural Score Estimation (FNSE) method adopts a "divide and conquer" strategy. Instead of learning from full trajectories $x^{0:T}$, the model is trained on individual transitions $(x^t, x^{t+1})$.

The FNSE framework: Training on single transitions allows for efficient local score estimation, which is then aggregated for global posterior inference.

The framework consists of three key components:

  1. Score Factorization: The global score (the gradient of the log-posterior) is decomposed into a sum of local scores corresponding to each transition in the sequence.
  2. Local Training: Neural networks are trained to estimate these local scores using pairs of states. This approach is highly efficient as it avoids full-sequence simulations and allows for targeted sampling via a proposal distribution.
  3. Global Composition: During inference, local score estimates are aggregated across the observed time series using composition rules to reconstruct the full posterior.

Experiments and Results

We evaluated FNSE across a range of benchmarks, including synthetic SDEs and classical scientific models like Lotka-Volterra and SIR. A significant challenge was the Kolmogorov flow simulator, which models incompressible fluid dynamics with approximately one million data dimensions per time series.

Benchmark results showing the sliced Wasserstein distance (sW1) for various tasks. FNSE and related factorized methods (FNLE, FNRE) consistently outperform standard NPE.

Key results include:

  • Simulation Efficiency: FNSE and its variants (FNLE, FNRE) consistently outperformed standard Neural Posterior Estimation (NPE), requiring fewer simulator calls to reach higher accuracy.
  • Scalability: FNSE successfully performed inference on the million-dimensional Kolmogorov flow task using only 200k simulator steps, demonstrating a capability beyond standard SBI methods.
  • Calibration and Sampling: The method produced well-calibrated uncertainty estimates and enabled fast posterior sampling via diffusion-based techniques, which is particularly beneficial for long time series.

Conclusion

By leveraging the Markovian structure of time series data, FNSE provides a scalable and efficient framework for simulation-based inference. While the method's performance depends on the training proposal distribution and the Markovian assumption, its ability to handle high-dimensional scientific models marks a significant advancement in the field. Future research will explore extensions to non-Markovian systems and integration with data assimilation techniques.

Summary updated

BibTeX
@inproceedings{glockler2025compositional,
  author = {Manuel Glöckler and Shoji Toyota and Kenji Fukumizu and Jakob H. Macke},
  title = {{Compositional simulation-based inference for time series}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=uClUUJk05H},
}
Figure