Terra Laboratories
The science of inference. The business of keeping things alive.
Understand the problem01 The problem
Every instrument in a tank, a digester or an aeration basin measures a consequence. Dissolved oxygen is what remains after the biology has taken what it required. Ammonia is what the biofilter has yet to clear. The variable that determines the outcome carries no instrument at any price, so operators fund standing headroom and night shifts respond to symptoms.
02 The science
The biology is expressed as a mechanistic model. That model is solved against the instrumentation already installed in the process to recover the unmeasured state, then integrated forward to produce a horizon. The output is a probability of failure within a stated window, with a confidence interval, rather than another reading.
See it applied
03 Flagship
A sealed habitat is a recirculating farm without buffer or exit. Every by-product accumulates inside the loop, and the interval between stable and uninhabitable is measured in hours.
Terra was validated against SIMOC across 60 runs and 7 configurations, including Biosphere 2 mission configurations and a Mars habitat model, under benchmark TE/06. Oxygen breaches were called 64 to 111 hours ahead of the physical gauge. Terrestrial systems present the same problem with materially more tolerance.
The frontier04 System index
Recirculating aquaculture, hydroponics, aquaponics, bioremediation, biomanufacturing and closed-loop life support. Six industries with separate trade shows and separate vocabularies, running the same physics. Each is a functional block of a bioregenerative life support system (BLSS), and each already exists on Earth with an operating budget, which is where the engine is built and proven before it is closed into a loop.
01BacktestedBiofilter capacity, the oxygen budget, and the position of the nitrogen cycle, on a recirculating farm where the whole loop turns over several times an hour.
6.5 hours of usable lead reproduced on held-out commercial operating history.S8
02Validation openThe physiological state of a crop, inferred from the environment it actually experienced rather than from the set points it was supposed to receive.
Retrospective test in design with operators. Result published on completion of the run.
03Validation openTwo biologies sharing one water column, where the nitrogen the fish produce is the nitrogen the crop consumes and neither side can be tuned without moving the other.
Runs on the same state estimator as systems 01 and 02, coupled. Benchmark scoping with operators.
04ValidatedNitrifier activity and the oxygen demand running ahead of the blower response, in the reactors that carry the treatment load.
IWA BSM1 benchmark. Dissolved oxygen prediction error cut by 92 percent.S4,S8
05ValidatedMetabolic flux and the trajectory toward final yield, reconstructed between sparse and expensive assays.
E. coli dynamic flux balance. One-step error reduced by 82 percent.S6,S8
06ValidatedPartial pressures and the hours of margin to a life-critical bound, in a loop with no resupply and no exit.
SIMOC, 60 runs across 7 configurations. 98% of breaches identified before they occurred.S7,S8
05 Company
The compounding asset is validated domains. Each system Terra qualifies shortens the next, because the method transfers intact and only the biology is rewritten.
Two of the three modelling stages are invariant across domains. Only the biology specified in the first stage is domain-specific, which is what allows a single engine to cover this range.
Every figure carries a reference. Where a system is still in test, its entry states the stage it has reached rather than carrying a figure across from an adjacent domain.
False alarm rates are published alongside lead times, test stages are stated explicitly, and the conditions under which the method degrades are documented in advance of deployment.
06 Validated against
Terra is evaluated against the reference models each field already accepts, with independent laboratory validation running in parallel. These are the standards a process engineer would select to be judged by. Operator evaluations are open.
The international standard model for activated sludge treatment.
The reference model for anaerobic digestion processes.
Standard method for modelling microbial metabolism over time.
Agent-based closed life support simulator from Arizona State University, including Biosphere 2 mission configurations.
Terra's biofilter-health estimates checked against laboratory assay at an established recirculating aquaculture research facility.
07 Validation record
Four benchmark validations complete, one held-out backtest on commercial operating history, one independent laboratory validation in progress, and two domains in test design with operators. Every figure below is reproducible against a published model or a dated dataset.
Scope: every entry above is a benchmark validation or a retrospective backtest. Live process results are published to this same record as operator evaluations complete.
08 The frontier
Terra is building the building blocks of biological life support. A fish farm and a sealed habitat fail the same way, and the difference between them is buffer size and consequence, not physics. Earth funds the work. Every instrument deployed on the way makes the harder version better, and the failure trajectories that accumulate cannot be bought anywhere.
09 Common questions
Anomaly detection learns a picture of normal operation and flags departures from it, which tells an operator that something looks unusual at this moment. Terra estimates the physical state of the process and simulates that state forward, which is where the time horizon comes from. The two methods also have different data requirements. Anomaly detection needs a long history of labelled failures; state estimation needs a model of the process, which the field already has.
Most of it, and each figure is labelled with its source. The simulators are the reference models the relevant fields already use to evaluate control strategies: BSM1 for activated sludge, ADM1 for anaerobic digestion, dynamic flux balance for metabolism, and SIMOC for closed life support, which includes Biosphere 2 mission configurations. The aquaculture result is different in kind: it is a held-out backtest on real commercial operating history. The remaining work is commercial rather than scientific.
The dataset. Real failure trajectories are not available for purchase and accumulate only where an instrument is already deployed, so the advantage compounds with each operator. Beneath that, mechanistic process modelling is an uncommon skill set. And once a control room runs a night shift against a forecast, that forecast becomes part of the operating procedure rather than a tool sitting beside it.
A recirculating farm is a closed loop with material value at risk every night, which makes it the same physical problem as a habitat with a larger buffer and a far shorter sales cycle. Validating at the hardest end and commercialising at the most reachable one is the deliberate sequence.
An operator sends a historical export. Terra runs it blind with the outcome window withheld, then reports the hours of warning the engine would have given and how often it would have been wrong on that same history. There is no connection to any live system, no change to any control loop, and no fee. The file is deleted afterwards.
10 Evaluation
Provide a historical export. Terra runs it blind with the outcome window withheld and returns the hours of warning the engine would have given, together with its false alarm rate on that same history. No connection to any live system, no change to any control loop, and no fee. The file is deleted on completion.
11 Sources
Every figure on this page carries a reference mark. Industry figures resolve to the published record. Terra results resolve to the validation record in section 07, where run counts, configurations and test conditions are stated in full.
170,000 salmon lost overnight at a land-based farm in Japan after circulation pumps stopped and oxygen fell in two tanks. Reported loss NOK 12 million, about USD 1.16 million.
The Fish Site / SalmonBusiness, June 2025
Open source100,000 Atlantic salmon lost at a land-based farm in Nova Scotia after a carbon dioxide filter failed structurally. Reported loss USD 5 million, about 20 percent of stock on site.
WeAreAquaculture / CBC News, November 2023
Open sourceAeration accounts for 25 to 60 percent of energy use across wastewater treatment plants generally, and 50 to 90 percent of plant electricity in conventional activated sludge systems specifically.
University of Michigan Center for Sustainable Systems, US Wastewater Treatment Factsheet; Drewnowski et al., Processes 7(5):311, 2019
Open sourceThe international reference model for benchmarking control strategies on activated sludge plants.
Alex, Benedetti, Copp, Gernaey, Jeppsson, Nopens, Pons, Steyer and Vanrolleghem. IWA Task Group on Benchmarking of Control Strategies for WWTPs
Open sourceThe reference structured model for anaerobic digestion processes.
Batstone, Keller, Angelidaki, Kalyuzhnyi, Pavlostathis, Rozzi, Sanders, Siegrist and Vavilin. Water Science and Technology 45(10):65-73, 2002
Open sourceThe standard method for modelling microbial metabolism through time, introduced on diauxic growth in E. coli.
Mahadevan, Edwards and Doyle. Biophysical Journal 83(3):1331-1340, 2002
Open sourceAn agent-based simulator of closed life support systems, developed at Arizona State University's School of Earth and Space Exploration and hosted by the National Geographic Society. Configurations include Biosphere 2 mission parameters.
Arizona State University Interplanetary Initiative
Open sourceEvery Terra figure on this page resolves to the validation record in section 07, where run counts, configurations, error reductions, lead times and false alarm rates are stated in full. Simulation results are labelled as simulation. All results to date are benchmark validation and retrospective backtesting. Live process results are published to the same record as operator evaluations complete.
Terra Laboratories, benchmarks TE/01 and TE/04 through TE/06