As machine learning plays an increasingly important role in ocean science, CBIOMES researchers met in New York and online to advance shared tools for modeling the ocean’s dynamic biological systems.
Reporting by Helen Hill for CBIOMES
The CBIOMES Workshop on Machine Learning for Time‑Series Data brought together researchers from across the collaboration to explore modern computational approaches for analyzing ecological and biogeochemical time‑series data. Held at the Simons Foundation headquarters in New York City, the three‑day meeting combined lectures, hands‑on tutorials, and collaborative hackathon sessions designed to bridge machine-learning methods with traditional statistical modeling and mechanistic, process‑based approaches.
Time‑series data are central to CBIOMES science—from phytoplankton bloom dynamics to microbial community shifts and coupled physical–biogeochemical processes. The workshop focused on how machine‑learning methods such as vector autoregressive (VAR) models, sparse VAR, artificial neural networks (ANNs), and physics‑informed neural networks (PINNs) can be used to extract insight from these complex systems. Participants learned how to select, implement, and interpret appropriate methods, and how to build reproducible workflows using real‑world datasets, including those accessible through the Simons CMAP platform.
A Program Designed for Depth and Collaboration
The workshop was organized by Kevin Egan, Brian Powell, Jann Paul Mattern, Gregory Britten, Christian Müller, Chris Edwards, Leander Schwarzmeier, and Viet Tran, who crafted a program that blended conceptual lectures, hands‑on tutorials, and collaborative problem‑solving.
Participants began with foundational methods—vector autoregressive (VAR) models and their sparse extensions—before moving into nonlinear neural‑network architectures and physics‑informed neural networks (PINNs). Each session was paired with practical exercises using a shared Docker‑based environment and real‑world datasets accessed through the Simons CMAP platform. This reproducible workflow ensured that participants could continue building on the material long after the workshop ended.
Remote attendees joined seamlessly via Zoom, contributing to discussions and participating in tutorials alongside those in the room.
Hands‑On Learning with Real Data
A defining feature of the workshop was its emphasis on applied learning. Tutorials guided participants through:
-
- Constructing and interpreting VAR and sparse VAR models
- Training neural networks for ecological time‑series prediction
- Embedding mechanistic constraints into PINNs
- Accessing and processing global ocean datasets via CMAP
- Building reproducible workflows using containers and shared code
These sessions were designed not only to teach methods, but to help participants understand when and why to use them—balancing predictive performance, interpretability, and scientific relevance.
Community Building Beyond the Classroom
The workshop also strengthened the social fabric of the CBIOMES collaboration. Group dinners at Casa Carmen and Portale provided opportunities for informal discussion, idea‑sharing, and new collaborations. Conversations ranged from technical debates about regularization strategies to broader reflections on the future of hybrid modeling in ocean science.
A Hackathon to Synthesize and Innovate
The final day featured a collaborative hackathon, where participants applied the full suite of methods to their own datasets and research questions. Teams explored bloom dynamics, microbial interactions, and physical–biogeochemical coupling, experimenting with both data‑driven and mechanistically constrained models. One group began developing a physics-informed neural network based on the classic 1946 phytoplankton growth model of Gordon Riley, demonstrating how historical ecological theory can be integrated with modern machine-learning approaches. The atmosphere was energetic and collegial, with participants sharing code, troubleshooting together, and celebrating breakthroughs.
Impact and Future Directions
The workshop marked a significant step forward for the collaboration. Participants left with:
-
- A stronger foundation in modern machine‑learning methods
- Reproducible workflows that can be shared across institutions
- New collaborations and cross‑disciplinary connections
- A clearer sense of how machine learning can complement mechanistic modeling in marine science
Looking ahead, participants expressed enthusiasm for deeper dives into uncertainty quantification, hybrid mechanistic–ML modeling, and large‑scale ecosystem prediction. The workshop laid the groundwork for these next steps, reinforcing CBIOMES’ commitment to advancing computational approaches that illuminate the ocean’s complex and dynamic systems.
“One of the goals of the workshop was to help researchers better understand not only how to implement these methods, but when they are most appropriate for different scientific problems,” said Kevin Egan. “It was exciting to see participants engaging with everything from sparse statistical models to neural networks and physics-informed approaches, and thinking carefully about how these tools can complement mechanistic modeling in ocean science.”
“The most rewarding aspect of this workshop is that the majority of organizers, presenters, and attendees were young scientists,” said Brian Powell. “ The presenters, Leander, Viet, Kevin, and Paul, are all young investigators who assembled incredible teaching notebooks on how these advanced methodologies work and provided implementation examples. The participants were able to try these methods on their own data, and I am most enthusiastic that graduate students and young postdoctoral scholars will be able to advance their studies using these methods. It is rewarding to see the new generation of scientists leading these efforts.”
Special thanks to the meeting organizers and, of course, to the generosity of the Simons Foundation for underwriting this activity.
Participants
Attendees represented a broad cross‑section of the CBIOMES community. Participants included:
Olivia Ahern (MBL), Mohammad Amirian Matlob (Dalhousie), Mohammad Ashkezari (UW), Jacob Bien (USC), Gregory Britten (WHOI), Louis Christie (USC), Joost de Vries (Bristol), Chris Edwards (UCSC), Kevin Egan (UH), Mick Follows (MIT), Helen Hill (MIT), Alex Hochroth (WHOI), Sangwon Hyun (UCSC), Alexandra Jones‑Kellett (UH), Bror Jönsson (UNH), Žarko Kovac (University of Split), Oh‑Ran Kwon (Ohio State), Danling Ma (WHOI), Jann Paul Mattern (UCSC), Brian Powell (UH), Leander Schwarzmeier (Helmholtz Munich), Yuanyuan Song (MIT), Rose Terner (Dalhousie), Viet Tran (Helmholtz Munich), Joseph Vallino (MBL), Ziyue Zheng (UCSC).
Story Image: Attendees of the CBIOMES Workshop on Machine Learning for Time‑Series Data, NY April 2026 – image credit: H. Hill
April 2026 CBIOMES Workshop on Machine Learning for Time‑Series Data: Plenary recordings, agenda, attendee list (restricted)