Geobyte: Geology × Computing Lab
About Geobyte
Geobyte is the research group of Dr. Haipeng Li at the International Research Center for Paleogeography, Chengdu University of Technology. We work where sedimentary geology meets computing, building new ways to reconstruct the deep-time Earth.
We combine outcrops, cores, well logs, seismic data and the published literature with plate-tectonic models, knowledge graphs, machine learning and numerical simulation. The goal is not to replace geological judgement, but to make interpretation explicit, reproducible, and honest about non-uniqueness.
Our work follows the full path from a rock record to a paleogeographic map: observe, encode, interpret, reconstruct, and share.
We design systems that stay transparent about uncertainty, preserve the provenance of every observation, and help geologists ask better questions of the deep-time record.
Interpretation under uncertainty
Geological records are incomplete, and the same deposit can often be explained by more than one environment. We treat that non-uniqueness as something to measure rather than hide: information theory and Bayesian inference turn paleoenvironmental interpretation into a reproducible procedure (Li & Plink-Björklund, 2019, Geophysical Research Letters).
The same idea guides our work with large language models. AI tools should widen and rank candidate interpretations, in the spirit of Chamberlin’s multiple working hypotheses, rather than collapse ambiguity into one confident answer.
This thread began in the field: how environmental signals are propagated, preserved and finally identified in sediment routing systems, and what Eocene river deposits of the Green River Formation (Uinta Basin, Utah) record about supercritical flow.
We build for real research conditions: sparse data, competing hypotheses, multiple scales, and conclusions that have to be explained to collaborators.
Research directions
Our research directions share one question: how can computation make paleogeographic reconstruction quantitative, reproducible and open, while respecting the spatial, temporal and physical context of geological evidence?
Digital paleogeography. We lead TimeMachine, the 4D paleogeographic reconstruction platform of the International Association for Paleogeography (IAP): browser-based, GPU-accelerated, and connected to plate models and community databases such as PBDB, Macrostrat and GeoLexicon.
Knowledge-guided GeoAI. PaleolithoSystem links well logs, seismic, core photos and thin sections on one depth axis, and uses a sedimentary-facies knowledge graph to guide human–AI annotation. Expert interpretation, including alternative hypotheses and confidence, becomes provenance-tracked, training-ready data.
Seismic facies, framework first. Before classifying anything we build the stratigraphic framework: faults, a relative-geologic-time field and layer-bounded units. Reflection configurations are then classified with foundation-model features, structured geological reasoning, and an expert in the loop.
A forward-modelling workshop generates synthetic sections with exact ground truth. Key experiments are pre-registered, run with placebo controls, and logged in a ledger of verdicts and retractions, negative results included.
Surface processes. Building on a PhD on environmental signals in sediment routing systems, and an NSFC Young Scientists Fund project on the Yarlung River source-to-sink system, we use numerical models such as Landlab to ask which signals a river network can transmit and which its deposits can preserve.
The Geobyte team
Geobyte is a small group led by Dr. Haipeng Li (李海鹏), Associate Research Fellow and Assistant Director of the International Research Center for Paleogeography at Chengdu University of Technology, working with graduate students and research software collaborators.
“A good lab is a place where a geological intuition can become a prototype before it becomes a paper.”
“A good lab is a place where a geological intuition can become a prototype before it becomes a paper.”
Haipeng received his PhD in Geology from Colorado School of Mines in 2020 (advisor: Piret Plink-Björklund), after an MSc at China University of Geosciences (Beijing) and a BSc at China University of Petroleum (Beijing). He was a postdoctoral researcher at the Deep-time Digital Earth (DDE) Research Center of Excellence in Suzhou (2021–2023) and deputy director of the Global Paleogeography Research Center at the Zhejiang DDE international research center (2024–2025), and joined CDUT in 2025. He serves as Secretary-General of the International Association for Paleogeography and co-leads the DDE Paleogeography Working Group.
Platforms and projects
Geobyte projects turn deep-time evidence into shared instruments for reconstruction. TimeMachine (docs.deeptime.world) supports the Deep-time Digital Earth programme and serves TopoAsia, the international programme on the 4D topographic evolution of Asia approved by the International Lithosphere Program, where Haipeng is executive convener of the session on 4D paleogeographic reconstruction of Pan East and Southeast Asia.
Current work spans the Late Mesozoic paleogeography of East Asia, the comparison and evaluation of plate-tectonic models, AI-assisted lithofacies-paleogeographic mapping with industry partners, and a deep-time paleogeography subset for high-quality AI datasets.
The practical work matters as much as the model: cleaning data, designing experiments, checking assumptions, and building visual tools that make results easier to discuss.
Reliable geoscience
Geoscience decisions often carry incomplete observations and asymmetric consequences. We therefore treat reliability as a research problem, not a finishing step.
Our evaluations ask how models behave when interpreters disagree, labels are sparse, or the data come from outside the area used for training.
We compare predictions with independent measurements, document failure modes, and make uncertainty visible alongside every map, classification, and simulation.
A model that cannot explain what it saw, what it assumed, and where it may fail is not ready for the field.
We also test the human workflow around a tool: whether collaborators can reproduce a result, challenge it, and update it when new evidence arrives.
These practices help us build systems that are useful without overstating what the data can support.
We pre-register key experiments, keep a ledger of verdicts and retractions, and report negative results, so that reliability can be checked rather than simply claimed.
Responsibility also means respecting data owners: partner data stay with the partner, and our public examples use open benchmarks or generalized cases.
Open and responsible science
Geobyte practices open and responsible science: we share methods, record decisions, and invite domain experts to challenge our assumptions.
Sensitive datasets require care. We work with data owners and local partners to define appropriate access, attribution, and resolution before a project begins.
We prefer interpretable baselines before complex models, and we keep human review where an automated conclusion could shape a high-stakes decision.
Clear communication is part of the method. We distinguish observation, inference, simulation, and speculation in every public result.
Our software is designed for extension: new data types, new basins and new hypotheses should be able to join the system without rewriting its foundations.
We welcome collaborations that make Earth science more reproducible, more legible, and more connected to the communities it serves.
We are especially interested in partnerships that join careful sedimentology and stratigraphy with ambitious computational experimentation.
Join the lab
Geobyte welcomes geologists, computer scientists and engineers who want to work across disciplines. We value careful questions, generous collaboration, and tools that other researchers can actually use.
We work with graduate students at Chengdu University of Technology and welcome visiting researchers, student projects and engineering partnerships around digital paleogeography, GeoAI for sedimentary geology, and source-to-sink modelling.
To start a conversation, write to haipeng.li@cdut.edu.cn with a short note about the Earth question you want to explore and the perspective you bring.
A good way to meet us is the hands-on short course “Reconstruct Asia in 4D with TimeMachine” on 6 November 2026 in Beijing, held within the 2026 International Symposium on Deep Earth Exploration and TopoAsia.
Core team
Member | Discipline | Role | Contribution | Since | Contact |
Haipeng Li (李海鹏) | Sedimentology · paleogeography | PI · Associate Research Fellow | TimeMachine, GeoAI, source-to-sink | 2025 | haipeng.li@cdut.edu.cn |
Graduate researchers | Geology · geoinformatics | MSc students | Plate-model evaluation, source-to-sink modelling, facies annotation | 2025 | via the PI |
Research software collaborators | Web GIS · GPU computing | Engineering | TimeMachine, WebGPlates | Ongoing | via the PI |
Open position | Geology × computing | Student or visiting researcher | Your project | 2026 | haipeng.li@cdut.edu.cn |
Research directions
Area | Platform | Data | Question | Status | Lead |
Digital paleogeography | TimeMachine | Plate models + community databases | Where was it, and when? | Active | H. Li |
Knowledge-guided annotation | PaleolithoSystem | Logs, seismic, core, thin sections | How does expert knowledge become training data? | Released 2026 | H. Li |
Seismic facies | In-house prototype | Public 3D seismic benchmarks | Framework first, then facies | Research | H. Li |
Interpretation and LLMs | Facies knowledge graph | Literature + facies models | How non-unique is an interpretation? | Active | H. Li |
Source-to-sink modelling | Landlab | River networks, discharge | Which signals survive? | Active | H. Li |
Software and tools
Tool | Stack | Purpose | Licence | Version | Link |
TimeMachine | Web · GPU rasterization | 4D paleogeographic reconstruction | Online platform | Live | |
TimeMachine raster-assignment artifact | Code + data | GPU assignment of points to tectonic elements | MIT | — | |
PaleolithoSystem | FastAPI · React · SQLite · Neo4j | Knowledge-graph-guided multimodal annotation | All rights reserved | 2026 release | |
Deeptime Harvester | Zotero plugin | Label figures and extract knowledge from the literature | — | v0.5 beta | — |
Paleo-elevation 3D visualization system | — | Visualize paleo-elevation reconstructions | Software copyright (2023) | V1.0 | — |
Selected papers
Title | Year | Venue | Topic | Authors | Link |
Applying information theory and Bayesian inference to paleoenvironmental interpretation | 2019 | Geophysical Research Letters 46(24) | Interpretation | Li*, Plink-Björklund | |
板块构造模型对比与评估方法的研究进展 | 2026 | 高校地质学报 (online first) | Plate-model evaluation (review, in Chinese) | Yang, Li*, Hou, Cheng | |
沉积源—汇系统数值模拟研究进展：多模型比较与应用 | 2024 | 地球科学进展 39(11) | Source-to-sink modelling (review, in Chinese) | He, Li*, Hou | |
Using PySpark to accelerate batch data point rotation for paleogeographic reconstruction | 2024 | International Journal of Digital Earth 17(1) | Paleo-coordinates | Xu, Hu, Li, Qin, Wu, Du | |
Online data service for geologic formations (Lexicons) of China, India, Vietnam and Thailand with one-click visualizations onto East Asia plate reconstructions | 2024 | Geoscience Data Journal 11(4) | Formation databases | Du, Mishra, Ogg, … Li, Scotese, Dong | |
The progress and perspective of digital intelligence in comprehensive paleogeographic reconstruction (in Chinese) | 2023 | Acta Geologica Sinica 97(9) | Digital paleogeography | Hou, Chen, Ren, … Li, et al. | |
Hydrocarbon exploration potential of the Jurassic Chaoshan Subbasin in northern South China Sea: evidence from the latest seismic and outcrop data | 2023 | Geofluids | Basin analysis | Qiang, Li* | |
Environmental signal propagation, preservation, and identification in sediment routing systems | 2020 | PhD thesis, Colorado School of Mines | Source-to-sink | Li |
Active projects
Project | Timeline | Method | Role | Status | Region |
Environmental signal propagation and preservation in the Yarlung River source-to-sink system (NSFC Young Scientists Fund, 42302133) | 2024–2026 | Numerical modelling | PI | Active | Tibetan Plateau |
Late Mesozoic paleogeographic reconstruction of East Asia (sub-task, National S&T Major Project on Deep Earth) | 2024–2028 | Plate models + geological data | Sub-task lead | Active | East Asia |
Digital-intelligent paleogeographic reconstruction platform for TopoAsia and resource exploration (Sichuan international S&T cooperation programme) | 2026–2027 | TimeMachine | PI | Active | Asia |
Multimodal annotation technology and tools for intelligent lithofacies-paleogeographic reconstruction (CDUT AI + Science programme, 2025AI001) | 2025–2026 | Knowledge-graph-guided annotation | PI | Concluding | Continental lake basin |
AI-assisted lithofacies-paleogeographic mapping with an industry partner | 2025–2027 | Knowledge + data dual-driven | Core member | Active | Offshore rift basin |
Deep-time paleogeography subset, CDUT high-quality AI dataset initiative | 2026–2028 | Data curation + knowledge graph | Subset lead | Building | Global to basin |
Networks and partners
Partner | Type | Collaboration | Since | Scope | Status |
International Association for Paleogeography (IAP) | International association | Secretariat hosted at CDUT; TimeMachine platform | 2025 | Global | Active |
Deep-time Digital Earth (DDE) | Big-science programme | Paleogeography Working Group co-lead | 2024 | Global | Active |
TopoAsia | International programme (ILP) | Executive convener; synthesis tooling with TimeMachine | 2025 | Asia | Active |
Colorado School of Mines | Academic | Fluvial sedimentology, supercritical flow | 2015 | Uinta Basin, USA | Ongoing |
Industry partners | Industry | AI-assisted lithofacies mapping | 2025 | Basin scale | Active |
Lab values
Principle | Meaning | Practice | Shared | Priority | No. |
Reproducibility | Every result has a path to evidence | Protocol | Shared | Core | 01 |
Uncertainty | Confidence is shown, not implied | Model card | Shared | Core | 02 |
Domain expertise | Experts stay in the loop | Review | Shared | Core | 03 |
Open practice | Methods travel farther when shared | Repository | Shared | Core | 04 |
Data resources
Resource | Source | Format | State | Access | Action |
TimeMachine reconstructions | Plate models + PBDB / Macrostrat / GeoLexicon | Web maps | Online | Public | |
Deep-time database | Fossil occurrences, geochronology, reconstruction observations | PostgreSQL / PostGIS | In development | Internal | — |
Sedimentary-facies knowledge graph | Expert-reviewed literature extraction | SQLite + Neo4j | Growing | Open release planned | — |
Demonstration annotation dataset | Wells, 3D seismic, core photos | Training-ready packages | Curated | Restricted | Request |
Milestones
Milestone | Year | What changed | State | No. | Record |
IAP secretariat hosted at CDUT | 2025 | Memorandum signed on 24 October 2025 | Done | 01 | — |
TopoAsia launch | 2025 | Launch symposium and Shanghai Declaration, October 2025 | Done | 02 | |
PaleolithoSystem release | 2026 | Published on the IAP community platform, 10 July 2026 | Done | 03 | |
TimeMachine short course | 2026 | Reconstruct Asia in 4D, Beijing, 6 November 2026 | Upcoming | 04 | Join |
Portfolio snapshot, September 2026. Statuses describe the current research stage; partner data and unpublished results are not shown.
RESEARCH NOTES
- 1
Geobyte datasets retain measurement provenance, coordinate reference systems, and collection context whenever those details are available.
- 2
Model outputs are reviewed against held-out observations and domain-expert annotations; a high score is never treated as geological proof.
- 3
A paleogeographic reconstruction depends on the plate model behind it. We state the model used and, where possible, compare alternatives.
- 4
Simulation results describe a tested scenario and its assumptions. They are not forecasts without uncertainty and independent validation.
- 5
Most of our software is still in-house or in beta. Each release states its licence, its known limitations, and how to reproduce the published examples.
- 6
Industry and data partners retain control over their data and permissions. Public examples use open benchmarks or are generalized where necessary.
- 7
The team welcomes corrections, alternative interpretations, and requests for collaboration through haipeng.li@cdut.edu.cn.