applicationContext = Production

News

13 July, SFB1258 Colloquy with Christian Glaser und Kristian Tchiorniy

Please be invited to our fourth SFB1258 Colloquy in the summer semester 2026.

Program

9:30 Welcome Coffee

10:00 Christian Glaser (TU Dortmund): How Deep Learning and Differential Programming Accelerate Progress in Neutrino Astronomy

Abstract: In this talk, I will discuss how modern AI methods are accelerating progress in fundamental physics. Using examples from my research in neutrino astronomy, I will highlight recent methodological advances with broad applicability across physics. In particular, I will focus on three developments:
•End-to-end detector optimization:
Differentiable programming enables the construction of fully differentiable simulation and reconstruction pipelines, allowing for gradient-based optimization of high-dimensional detector design spaces, an approach that was previously infeasible. This is achieved by combining explicit differential programming, generative machine learning models, neural network surrogates, and uncertainty quantification based on Fisher information.
•Real-time deep learning–based triggering on embedded systems:
Neural-network-based triggers can significantly enhance neutrino detection rates, achieving up to a factor of two improvement over classical approaches. I will discuss the challenges associated with processing continuous data streams exceeding 1 GS/s on low-power (~10 W) FPGA systems, and present our solutions for enabling low-latency inference under strict resource constraints.
•Neural posterior estimation for event reconstruction:
Accurate uncertainty quantification is often as critical as point estimation, yet remains challenging for standard neural networks. I will present a novel hybrid architecture that combines neural networks (e.g., Transformers) with conditional normalizing flows to infer full posterior distributions. As a case study, I will describe a new reconstruction method for the IceCube Neutrino Observatory to estimate neutrino arrival directions. This approach achieves unprecedented angular resolution while predicting the full Posterior and maintaining well-calibrated uncertainties.

10:45 Coffee break

11:15 Kristian Tchiorniy (TUM): Optical Neutrino Telescope Design Optimization

12:15 End

13 July, SFB1258 Colloquy with Christian Glaser und Kristian Tchiorniy

Please be invited to our fourth SFB1258 Colloquy in the summer semester 2026.

Program

9:30 Welcome Coffee

10:00 Christian Glaser (TU Dortmund): How Deep Learning and Differential Programming Accelerate Progress in Neutrino Astronomy

Abstract: In this talk, I will discuss how modern AI methods are accelerating progress in fundamental physics. Using examples from my research in neutrino astronomy, I will highlight recent methodological advances with broad applicability across physics. In particular, I will focus on three developments:
•End-to-end detector optimization:
Differentiable programming enables the construction of fully differentiable simulation and reconstruction pipelines, allowing for gradient-based optimization of high-dimensional detector design spaces, an approach that was previously infeasible. This is achieved by combining explicit differential programming, generative machine learning models, neural network surrogates, and uncertainty quantification based on Fisher information.
•Real-time deep learning–based triggering on embedded systems:
Neural-network-based triggers can significantly enhance neutrino detection rates, achieving up to a factor of two improvement over classical approaches. I will discuss the challenges associated with processing continuous data streams exceeding 1 GS/s on low-power (~10 W) FPGA systems, and present our solutions for enabling low-latency inference under strict resource constraints.
•Neural posterior estimation for event reconstruction:
Accurate uncertainty quantification is often as critical as point estimation, yet remains challenging for standard neural networks. I will present a novel hybrid architecture that combines neural networks (e.g., Transformers) with conditional normalizing flows to infer full posterior distributions. As a case study, I will describe a new reconstruction method for the IceCube Neutrino Observatory to estimate neutrino arrival directions. This approach achieves unprecedented angular resolution while predicting the full Posterior and maintaining well-calibrated uncertainties.

10:45 Coffee break

11:15 Kristian Tchiorniy (TUM): Optical Neutrino Telescope Design Optimization

12:15 End

13 July, SFB1258 Colloquy with Christian Glaser und Kristian Tchiorniy

Please be invited to our fourth SFB1258 Colloquy in the summer semester 2026.

Program

9:30 Welcome Coffee

10:00 Christian Glaser (TU Dortmund): How Deep Learning and Differential Programming Accelerate Progress in Neutrino Astronomy

Abstract: In this talk, I will discuss how modern AI methods are accelerating progress in fundamental physics. Using examples from my research in neutrino astronomy, I will highlight recent methodological advances with broad applicability across physics. In particular, I will focus on three developments:
•End-to-end detector optimization:
Differentiable programming enables the construction of fully differentiable simulation and reconstruction pipelines, allowing for gradient-based optimization of high-dimensional detector design spaces, an approach that was previously infeasible. This is achieved by combining explicit differential programming, generative machine learning models, neural network surrogates, and uncertainty quantification based on Fisher information.
•Real-time deep learning–based triggering on embedded systems:
Neural-network-based triggers can significantly enhance neutrino detection rates, achieving up to a factor of two improvement over classical approaches. I will discuss the challenges associated with processing continuous data streams exceeding 1 GS/s on low-power (~10 W) FPGA systems, and present our solutions for enabling low-latency inference under strict resource constraints.
•Neural posterior estimation for event reconstruction:
Accurate uncertainty quantification is often as critical as point estimation, yet remains challenging for standard neural networks. I will present a novel hybrid architecture that combines neural networks (e.g., Transformers) with conditional normalizing flows to infer full posterior distributions. As a case study, I will describe a new reconstruction method for the IceCube Neutrino Observatory to estimate neutrino arrival directions. This approach achieves unprecedented angular resolution while predicting the full Posterior and maintaining well-calibrated uncertainties.

10:45 Coffee break

11:15 Kristian Tchiorniy (TUM): Optical Neutrino Telescope Design Optimization

12:15 End

Back
Print