About

Ellen Marsh: engineer and statistician. PhD researcher at Imperial College London (Digital Twin for Bioprocess Design). Consulting in probabilistic forecasting and Bayesian modelling through Fedelm Advisory.

I’m Ellen Marsh. I trained as an engineer at Oxford, moved into statistics and machine learning through bioprocess modelling, and now divide my week between a PhD at Imperial College London and consulting through Fedelm Advisory. The consulting work is probabilistic forecasting and hierarchical Bayesian modelling for organisations that plan or allocate on the output: NHS trusts deciding on winter beds, R&D teams deciding which formulation to test next, commercial teams deciding where the next pound of spend goes.

The engineering training shows in how the work is delivered. A model leaves my hands with the baseline it had to beat, the coverage of its intervals measured on held-out data, a document the decision-maker can read in five minutes, and enough monitoring and handover material that it keeps running after the engagement ends. When a simpler method wins, I ship the simpler method and say so in the write-up.

What I work on

Forecasting and evaluation. Probabilistic forecasting with proper scoring rules; rolling-origin backtests; revision-aware evaluation on archived data vintages; conformal prediction intervals; hierarchical forecasting and reconciliation; gradient-boosted and classical models side by side. Scored, public forecasts of NHS England A&E activity from October 2026.

Bayesian modelling. Hierarchical models in PyMC; non-centred parameterisations; partial pooling and cold-start prediction for new units; simulation-based calibration for identifiability; prior and posterior predictive checks; CRPS and PIT diagnostics; Bayesian parameter inference from noisy time-course data, including a custom MCMC sampler built for an industrial bioreactor pipeline.

Bioprocess modelling and scale-up. Stirred-tank bioreactors and scale-up heterogeneity; CFD lifelines and the per-cell environmental histories they produce; surrogate models trained on them; hybrid mechanistic and data-driven process models. Constraint-based metabolic modelling: flux balance analysis, dynamic FBA, co-culture cross-feeding. Cultivated meat and cell-therapy media optimisation; precision fermentation. Bioreactor control: set-point tracking, disturbance rejection, model-predictive control, and the gap between the textbook and a real vessel with sensor noise and actuator lag.

Experimental design and optimisation. Design of experiments driven by parameter uncertainty; Bayesian optimisation; active and transfer learning across campaigns.

Techno-economics. Techno-economic and life-cycle assessment for pilot-scale cultivated-meat processes; medium recycling; CO2 utilisation.

Expert calls

I take one-hour calls through expert networks and directly. Topics I cover well:

  1. Scale-up failure modes in alternative-protein fermentation and cultivated-cell processes: mixing, gradients, and what the cells actually experience.
  2. What bioprocess digital twins can and cannot do today, and how to tell a twin from a dashboard.
  3. Assessing modelling and CFD vendor claims: the questions that reveal whether a model has been validated.
  4. Media optimisation methods and their evidence: design of experiments, active learning, Bayesian optimisation, and what the published comparisons show.
  5. Forecasting and experimental-design methods for R&D and operations: baselines, calibration, and how many experiments a decision needs.

I discuss public knowledge and my own judgement. I do not disclose unpublished results or confidential material from any employer, client or collaborator, past or present.

Background

PhD, Imperial College London (start TBD–). Digital Twin for Bioprocess Design: model discovery from large, uneven datasets; prediction; Bayesian optimisation; active and transfer learning. [TBD: department, supervisor and funder if they may be named]

MRes Engineering Biology, University of Bristol (2025–26, first class). Two research projects, both hybrid bioprocess models. In the first I built a constraint-based model of an engineered E. coli strain designed to take up mammalian-cell metabolites in a circular medium-recycling scheme for cultivated meat, coupled it to a flux-balance framework to locate engineering targets and operating points for medium supply, and mapped the regime in which the supporting strain becomes the growth-limiting partner. In the second I coupled CFD to cell models for E. coli in a stirred tank and trained a machine-learning surrogate to predict the environmental history each cell experiences over a run, for biomass and growth prediction.

MEng Engineering Science, University of Oxford, Trinity College (2020–25, first class). Head of Department Award for Excellent Performance in 2024 and 2025; Gibbs Prize for the best performance in the Final Honour School in 2023 and 2024. Coursework in optimisation and control, robust and multivariable control, model-predictive control, machine learning and neural-network architectures, and bioprocess engineering. My master’s project designed, fabricated and validated electroactive scaffolds for tissue culture under electrical stimulation, with MATLAB pipelines to extract quantitative biological response from a large set of video recordings.

Data Scientist, Extracellular, Bristol (July–October 2025). I joined an alternative-protein company to build a cell model and use it to optimise the bioprocess. I wrote a Bayesian pipeline in Python, with a custom MCMC sampler, to infer growth parameters from noisy bioreactor time courses and a mechanistic cell model, so that every estimate carried its uncertainty; I used those uncertainties to design the next experiments, trading parameter precision against experimental cost; I worked on bioreactor control logic for set-point tracking and disturbance rejection, and learned where control theory and industrial tuning part ways under sensor noise, actuator lag and biological drift; I supported the cloud infrastructure for remote monitoring and control; and I built the data pipelines that combine multi-source experimental records reproducibly.

Alternative protein. Cellular Agriculture UK mentorship scheme, mentored by Professor Eirini Theodosiou, covering the technical, commercial and regulatory landscape of cultivated meat and biomass fermentation. A design study for a pilot-scale cultivated-meat process: cell selection, bioreactor design and operating points, scale-up economics, life-cycle and techno-economic assessment, medium recycling and CO2 utilisation.

Publication. A data-driven review of in vitro electrical and mechanical stimulation for post-acute-phase wound healing. Advanced Healthcare Materials, 2026. doi:10.1002/adhm.71138. I curated a structured dataset from more than 100 studies and fitted random-forest regressions with feature-importance analysis to relate stimulation parameters to cellular response, identifying the parameter ranges that drive healing outcomes. The datasets catalogue on this site applies the same approach, a published literature brought into one schema before any model is fitted, to cell-culture media.

Other work. A Python pipeline for preprocessing, quality control and batch analysis of noisy cancer-cell voltage recordings, with a robust spike-detection algorithm (2025). A literature and coding study of learned representations of metabolic network structure: graph neural networks, hybrid models and stochastic network models.

Tools

Python (PyMC, LightGBM, scikit-learn, PyTorch, COBRApy, pandas), MATLAB, Git and GitHub Actions. I ship models as repositories with tests, not as notebooks.

Fedelm Advisory

Fedelm Advisory is the company through which I consult. It takes its name from the seer in the Táin Bó Cúailnge who is asked, before the battle, what she sees. I answer the same question with a range and the evidence for it.

Contact

email address TBD. Code at github.com/ellenm1612. Profile on LinkedIn. [TBD: LinkedIn URL]