ARTIST, an open-source, PyTorch-based differentiable ray tracer for solar thermal power plants¶
Speaker: Max Pargmann, DLR
Abstract¶
Solar tower power plants concentrate sunlight from thousands of mirrors (heliostats) onto a central receiver. While they could deliver high-temperature heat for chemical processes and cheap, dispatchable energy, solar tower plants have not taken off. The main barrier is complexity: individual component errors undermine the high accuracy required, yet tight cost constraints rule out dedicated metrology hardware. Their design and operation decompose into many problems that have each traditionally demanded a bespoke method — heliostat field layout, kinematic calibration of tracking errors, reconstruction of mirror surfaces, or aim point optimization to shape the receiver flux.
We present ARTIST, an open-source, PyTorch-based differentiable ray tracer that recasts these problems as gradient-based optimization through a single physical forward model (Pargmann et al., 2024). The forward pass aligns each heliostat via a rigid-body kinematic model, represents mirror facets as differentiable NURBS surfaces, and traces sampled rays — including sun shape and environmental conditions — to the receiver and calibration targets, accumulating a predicted flux density distribution. Reverse-mode AD then propagates a loss back to whichever parameters carry gradients: NURBS control points (surface reconstruction), the kinematic deviation parameters per heliostat (calibration), heliostat ground positions (field layout), or per-heliostat aim points (flux homogenization to suppress receiver hotspots). The same primal serves both inverse tasks, where the loss measures mismatch to recorded flux images, and design tasks, where it encodes an efficiency or flux-quality objective.
The substance of this framework is in differentiating the tracer at scale: handling discontinuities at ray–surface intersections, controlling variance in gradients through Monte Carlo ray sampling, and keeping the reverse-mode tape tractable across thousands of heliostats via a parallelized NURBS implementation, GPU acceleration, and distributed computation. We validate against real operational data from the Jülich solar tower, released openly through the PAINT database (Phipps et al., 2026). The result is a unified digital-twin environment where adding a new task means choosing new parameters to differentiate, not building a new solver.
References¶
Pargmann, M., Ebert, J., Götz, M., Maldonado Quinto, D., Pitz-Paal, R., & Kesselheim, S. (2024). Automatic heliostat learning for in situ concentrating solar power plant metrology with differentiable ray tracing. Nature Communications, 15. https://doi.org/10.1038/s41467-024-51019-z
Phipps, K., Kuhl, M., Weiel, M., Busch, M., Lewen, J., Blumenröhr, N., Maldonado Quinto, D., Debus, C., Göhring, F., Streit, A., Pitz-Paal, R., Götz, M., & Pargmann, M. (2026). PAINT: The First FAIR Database for Concentrating Solar Power Plants. Nature Energy.
