Surfing

Short Bio

PhD student at University of California, Irvine (UCI), advised by Stephan Mandt. Expert programming knowledge with 10+ years experience across a diverse range of languages and tasks including a long history of successful machine learning projects. Research in the intersection of Computer Science and Mathematics, currently focused on Deep Generative Models, Neural Data Compression, and the application of Machine Learning to Climate Science.

  • October 2020 - December 2022
    Research Assistant, TU Kaiserslautern1, Machine Learning Group
    Conducted various research of use in chemical process engineering and beyond. Developed a new tensor completion framework to make predictions for sparse tabular data and style-transfer methods for time series. Ongoing collaboration, including as invited speaker at a Dagstuhl seminar.
  • October 2019 - May 2020
    Research Assistant, German Research Center for Artificial Intelligence (DFKI)
    Developed an evolutionary algorithm to optimize the topology and hyperparameters of convolutional networks. Designed a front and back end providing 50+ users intuitive access to the local GPU computation cluster.
  • September 2018 - Current
    Student / Teaching Assistant, TU Kaiserslautern1 / UC Irvine
    Supported 1000+ students across 10+ courses in various roles as supervisor, mentor, advisor, educator, and examiner. Topics include probability theory, statistics, scientific computing, programming, machine learning, and more.
  • 1 Since 2023: RPTU in Kaiserslautern
  • Python
    Expert [8+ years] − Machine Learning Research, Data Analysis, Visualization, Hackathons, Coding Competitions, Tool and Application Development, Educator. PyTorch, TensorFlow, Sklearn, Django, etc.
  • R
    Expert [3+ years] − Statistical Analysis, Data Analysis, Educator.
  • MATLAB
    Expert [3+ years] − Optimization, Numerical Methods, Scientific Computing, Educator.
  • Java
    Expert [5+ years] − App and Game Development, Distributed Computing, Algorithm Design, Educator.
  • C / C++
    Advanced [2+ years] − Software Development, Algorithms and Data Structures.
  • HTML / CSS / JavaScript
    Advanced [2+ years] − Full-Stack Development with Python/Django Back-End.
  • Others
    Git, CUDA, Slurm, Docker, Kubernetes, SQL, Google/Microsoft Office Suite, VBA, Latex, UML, AI Tools, etc.
  • German
    Fluent
  • English
    Fluent
  • French
    B2 / Advanced Mid − 7+ years of instruction, student exchange with Ermont, France.
  • Spanish
    A2 / Intermediate High − 2+ years of instruction.
  • Swedish
    A2 / Intermediate Mid − 1+ years of instruction, semester abroad at Lund University, Sweden.
  • Sports
    Snowboarding, Surfing, Volleyball, Football
  • Culture
    Travel, Cooking, Winemaking

Research Interests

Networks

Deep Generative Models

Variational Autoencoders, Diffusion Models, their applications, and how to improve their inference.

Save-file

Neural Data Compression

Focus on efficiency, practicality, and scalability - for example as neural progressive codecs.

Science

Machine Learning
and Science

Including Chemical Process Engineering and Climate Science, in particular Climate Modeling and Cloud Microphysics.

Publications

    • NeurIPS
    • Conference
    • Best Paper
    ClimSim: A Large Multi-Scale Dataset For Hybrid Physics-ML Climate Emulation
    Sungduk Yu, Walter Hannah, Liran Peng, Jerry Lin, Mohamed Aziz Bhouri, Ritwik Gupta, Björn Lütjens, Justus Will et al.
    Neural Information Processing Systems 2023 (Outstanding Paper Award; top 0.05% of submissions)
  2. Modern climate projections lack adequate spatial and temporal resolution due to computational constraints. A consequence is inaccurate and imprecise predictions of critical processes such as storms. Hybrid methods that combine physics with machine learning (ML) have introduced a new generation of higher fidelity climate simulators that can sidestep Moore’s Law by outsourcing compute-hungry, short, high-resolution simulations to ML emulators. However, this hybrid ML-physics simulation approach requires domain-specific treatment and has been inaccessible to ML experts because of lack of training data and relevant, easy-to-use workflows. We present ClimSim, the largest-ever dataset designed for hybrid ML-physics research. It comprises multi-scale climate simulations, developed by a consortium of climate scientists and ML researchers. It consists of 5.7 billion pairs of multivariate input and output vectors that isolate the influence of locally-nested, high-resolution, high-fidelity physics on a host climate simulator’s macro-scale physical state. The dataset is global in coverage, spans multiple years at high sampling frequency, and is designed such that resulting emulators are compatible with downstream coupling into operational climate simulators. We implement a range of deterministic and stochastic regression baselines to highlight the ML challenges and their scoring.
    • AGU
    • Conference
    ClimSim-Online: A Large Multi-scale Dataset and Framework for Hybrid ML-physics Climate Emulation
    Sungduk Yu, Zeyuan Hu, Akshay Subramaniam, Walter Hannah, Liran Peng, Jerry Lin, Mohamed Aziz Bhouri, Ritwik Gupta, Björn Lütjens, Justus Will et al.
    AGU 2024
  4. Modern climate projections lack adequate spatial and temporal resolution due to computational constraints, leading to inaccuracies in representing critical processes like thunderstorms that occur on the sub-resolution scale. Hybrid methods combining physics with machine learning (ML) offer faster, higher fidelity climate simulations by outsourcing compute-hungry, high-resolution simulations to ML emulators. However, these hybrid ML-physics simulations require domain-specific data and workflows that have been inaccessible to many ML experts. As an extension of the ClimSim dataset (Yu et al., 2024), we present ClimSim-Online, which also includes an end-to-end workflow for developing hybrid ML-physics simulators. The ClimSim dataset includes 5.7 billion pairs of multivariate input/output vectors, capturing the influence of high-resolution, high-fidelity physics on a host climate simulator’s macro-scale state. The dataset is global and spans ten years at a high sampling frequency. We provide a cross-platform, containerized pipeline to integrate ML models into operational climate simulators for hybrid testing. We also implement various ML baselines, alongside a hybrid baseline simulator, to highlight the ML challenges of building stable, skillful emulators. The data and code are publicly released to support the development of hybrid ML-physics and high-fidelity climate simulations.
    • NeurIPS
    • Workshop
    Understanding and Visualizing Droplet Distributions in Simulations of Shallow Clouds
    Justus Will, Andrea Jenney, Kara Lamb, Michael Pritchard et al. and Stephan Mandt
    Machine Learning and the Physical Sciences Workshop - Neural Information Processing Systems 2023
  6. Thorough analysis of local droplet-level interactions is crucial to better understand the microphysical processes in clouds and their effect on the global climate. High-accuracy simulations of relevant droplet size distributions from Large Eddy Simulations (LES) of bin microphysics challenge current analysis techniques due to their high dimensionality involving three spatial dimensions, time, and a continuous range of droplet sizes. Utilizing the compact latent representations from Variational Autoencoders (VAEs), we produce novel and intuitive visualizations for the organization of droplet sizes and their evolution over time beyond what is possible with clustering techniques. This greatly improves interpretation and allows us to examine aerosol-cloud interactions by contrasting simulations with different aerosol concentrations. We find that the evolution of the droplet spectrum is similar across aerosol levels but occurs at different paces. This similarity suggests that precipitation initiation processes are alike despite variations in onset times.
    • ECML
    • Workshop
    Enhancing Realism in Batch Distillation Simulations: Data-Efficient Time Series Style Transfer with Transformers
    Justus Will, Justus Arweiler, Indra Jungjohann, Jennifer Werner, Mayank Nagda et al.
    Machine Learning for Chemistry and Chemical Engineering Workshop - European Conference on Machine Learning 2024
  8. In chemical process engineering, the accuracy and realism of simulated data are crucial for the effective design and optimization of a wide range of processes. In this paper, we demonstrate the efficacy of neural style transfer methods to enhance the realism of time series data generated from simulations. Specifically, this machine learning technique allows us to learn the style characteristics of non-parallel experimental data obtained from real-world chemical plants and then use them to transform simulated data to more closely reflect the realistic behaviors and variabilities not captured by the simulation model. We propose a transformer-based architecture with a latent representation that disentangles content and style information. After training, the underlying generative model allows for fast and data-efficient stylized generation without requiring many iterations of gradient-based optimization per sample, as in other time series style transfer baselines. We show the efficacy of our approach on both synthetic data and in an application to batch distillation.
    • NeurIPS
    • Workshop
    Towards Scalable Compression with Universally Quantized Diffusion Models
    Yibo Yang*, Justus Will*, Stephan Mandt
    Machine Learning and Compression Workshop - Neural Information Processing Systems 2024 (also submitted to ICLR 2024)
  10. Diffusion probabilistic models have achieved mainstream success in many generative modeling tasks, from image generation to inverse problem solving. A distinct feature of these models is that they correspond to deep hierarchical latent variable models optimizing a variational evidence lower bound (ELBO) on the data likelihood. Drawing on a basic connection between likelihood modeling and compression, we explore the potential of diffusion models for progressive coding, resulting in a sequence of bits that can be incrementally transmitted and decoded with progressively improving reconstruction quality. Unlike prior work based on Gaussian diffusion or conditional diffusion models, we propose a new form of diffusion model with uniform noise in the forward process, whose negative ELBO corresponds to the end-to-end compression cost using universal quantization. We obtain promising first results on image compression, achieving competitive ratedistortion-realism results on a wide range of bit-rates with a single model, bringing neural codecs a step closer to practical deployment.

Selected Projects

Project 1

Time Series Style Transfer

In applications to time series, style transfer can enhance realism in simulation data, improving performance on downstream tasks when high-quality training data is limited. Our transformer-based approach disentangles content and style in the latent space of a stylized autoencoder enabling fast and effective stylization. CODE MORE
Project 2

Detecting Faulty Concrete

Non-intrusive early detection of cracks in concrete is a crucial task in construction and maintenance. In collaboration with the Fraunhofer Institute for Industrial Mathematics (ITWM), we developed a computer vision framework that allows for fast and fully automated localization based on 3D CT scans, focusing on reliability and trustworthiness. CODE MORE
Project 3

Orthogonal Inductive Tensor Completion

Predicting missing entries in sparse tabular data is a key technique, for example in recommender systems. With the assumption that complex effects are dominated by interactions between small sets of explanatory variables, we can make efficient predictions for high-dimensional data from limited observations. CODE MORE
Project 4

ConvNEAT

Hyperparameter tuning of modern neural architectures is essential for final model performance and usually requires expert knowledge. We developed evolutionary algorithms that automatically find and tune viable architectures and demonstrate that they perform on par in image classification tasks. CODE MORE
Project 5

Walk Flow

In this Hackathon project, we developed tools that support effective planning of better pedestrian infrastructure and tourism services. To this end, we leverage machine learning and data analysis techniques to extract actionable insights from a large dataset of pedestrian flow. CODE MORE
Project 6

GPU Allocation UI

Based on Kubernetes and Docker, this repository provides users an intuitive graphical user interface to allocate, run, and prototype on the local GPU cluster. The web-based UI is fully integrated and accessible through the appropriate intranet endpoints. CODE MORE
View MoreLess Projects