EngD position: Designing a fast surrogate emulator for a microbiome simulator
Posted 7
views
Wageningen University & Research EngD position: Designing a fast surrogate emulator for a microbiome simulator
Description
Are you an engineer or applied data scientist who likes working with complex computational models? Are you curious about how soils store carbon and how microbes drive it? Then this EngD position in the Soil Biology Group might be for you!
Microbes decompose organic matter in soils and, in doing so, shape the carbon cycle. DEMENTpy (Decomposition Model of Enzymatic Traits) is a spatially explicit, trait- and individual-based simulator of this process. It represents enzymatic substrate degradation, microbial uptake, metabolism, mortality, reproduction, and dispersal across a spatial grid, which makes it one of the most realistic tools available for studying soil carbon cycling. That realism comes at a price: each simulation is computationally expensive, which limits the use of the model for sensitivity analysis, scenario exploration, and decision support.
In this project, you will design and build a validated statistical emulator of DEMENTpy: a fast surrogate model, trained on a systematic set of simulator runs, that reproduces DEMENTpy outputs, including estimates of uncertainty, at significantly reduced computational cost. You will evaluate constraints on generalizability of the model. You will use modern machine learning approaches for small data sets, such as the tabular foundation model TabPFN and Gaussian processes, and benchmark them against alternatives. You will deliver a documented Python package that allows soil ecologists to predict DEMENTpy outputs without having to run the full simulator.
YOUR ROLE
Link to the chairgroup:
Microbes decompose organic matter in soils and, in doing so, shape the carbon cycle. DEMENTpy (Decomposition Model of Enzymatic Traits) is a spatially explicit, trait- and individual-based simulator of this process. It represents enzymatic substrate degradation, microbial uptake, metabolism, mortality, reproduction, and dispersal across a spatial grid, which makes it one of the most realistic tools available for studying soil carbon cycling. That realism comes at a price: each simulation is computationally expensive, which limits the use of the model for sensitivity analysis, scenario exploration, and decision support.
In this project, you will design and build a validated statistical emulator of DEMENTpy: a fast surrogate model, trained on a systematic set of simulator runs, that reproduces DEMENTpy outputs, including estimates of uncertainty, at significantly reduced computational cost. You will evaluate constraints on generalizability of the model. You will use modern machine learning approaches for small data sets, such as the tabular foundation model TabPFN and Gaussian processes, and benchmark them against alternatives. You will deliver a documented Python package that allows soil ecologists to predict DEMENTpy outputs without having to run the full simulator.
YOUR ROLE
- use sampling methods to explore the parameter space of DEMENTpy systematically (initial microbial community composition and diversity, soil temperature, moisture, and litter chemistry such as C:N ratio and lignin content);
- carry out a sensitivity analysis (Morris elementary effects) to identify which parameters drive most of the variance in key outputs (litter mass loss, CO₂ flux, microbial biomass carbon, enzyme activity) and to focus simulation effort where it matters most;
- build the training data set by running DEMENTpy on a high-performance computing (HPC) system, summarizing spatial outputs per run and averaging stochastic replicates;
- train and validate emulators for each output variable (TabPFN regressors or Gaussian processes), and report clearly where emulator fidelity is sufficient;
- package the emulator as a lightweight, documented Python interface that accepts ecological inputs (litter quality and environmental variables like soil temperature and moisture) and returns predictions with calibrated confidence intervals, ready for use by soil scientists, microbial ecologists, and environmental scientists in research and education.The tool will ultimately be disseminated via a GitLab repository and an interactive web portal.
- follow the EngD training program and report on your design and results in a professional report and in presentations to ecologists and modelers.
Link to the chairgroup:
Skills
Sign in to read and join the discussion.