How The Models Work
Most agricultural AI learns correlations between weather, soils, and yield outcomes from large historical datasets — producing predictions that can be statistically calibrated but physiologically ungrounded. Aluvial takes a different approach. Our models are rooted in mechanistic plant science — we simulate the underlying biological and physical processes governing crop growth rather than relying solely on statistical pattern matching.
At the foundation of the platform is a biophysical crop model library calibrated against controlled-environment and field observations across corn, soybean, wheat, canola, camelina, pennycress, and broader oilseed systems.1Jones, J. W. et al. The DSSAT cropping system model. Eur. J. Agron. 18, 235–265 (2003). doi.org/10.1016/S1161-0301(02)00107-72Keating, B. A. et al. An overview of APSIM, a model designed for farming systems simulation. Eur. J. Agron. 18, 267–288 (2003). doi.org/10.1016/S1161-0301(02)00108-9 Weather forcing is derived from ERA5-Land,3Muñoz-Sabater, J. et al. ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data 13, 4349–4383 (2021). doi.org/10.5194/essd-13-4349-2021 PRISM, Daymet, and WorldClim, while soil constraints are parameterized using SSURGO, gSSURGO, and SoilGrids.4Hengl, T. et al. SoilGrids250m: Global gridded soil information based on machine learning. PLoS ONE 12, e0169748 (2017). doi.org/10.1371/journal.pone.0169748 This physiological foundation enables Aluvial to model Genotype × Environment × Management (G×E×M) interactions5van Ittersum, M. K. & Rabbinge, R. Concepts in production ecology for analysis and quantification of agricultural input-output combinations. Field Crops Res. 52, 197–208 (1997). doi.org/10.1016/S0378-4290(97)00037-3 directly and translate them into deployment intelligence, risk surfaces, and carbon intensity projections.
The underlying failure mode in agricultural commercialization is rarely product efficacy in isolation — it is the Commercial Gap: the disconnect between controlled R&D success and environmental variability under real-world deployment. A biological showing an average 8% relative yield response may still fail commercially if 30% of target environments fall below the economic response threshold due to soil water limitations, nitrogen availability, or adverse weather realization. Aluvial is designed to quantify that uncertainty before large-scale deployment decisions are made.
Background
A Brief History of Biophysical Crop Modeling
Biophysical crop simulation emerged from the Dutch school of production ecology in the late 1950s, with C. T. de Wit's foundational work on radiation interception, transpiration, and dry matter assimilation establishing the mechanistic basis for whole-plant growth modeling. Ritchie's soil water balance and evapotranspiration formulations6Ritchie, J. T. Model for predicting evaporation from a row crop with incomplete cover. Water Resour. Res. 8, 1204–1213 (1972). doi.org/10.1029/WR008i005p01204 provided the hydrological backbone that subsequent systems built upon.
By the 1980s, the CERES model family codified species-specific phenology and nitrogen cycling within a daily time step. These were integrated into the Decision Support System for Agrotechnology Transfer (DSSAT),1Jones, J. W. et al. The DSSAT cropping system model. Eur. J. Agron. 18, 235–265 (2003). doi.org/10.1016/S1161-0301(02)00107-7 enabling multi-season simulation of cropping sequences under variable management and weather. The Agricultural Production Systems sIMulator (APSIM)2Keating, B. A. et al. An overview of APSIM, a model designed for farming systems simulation. Eur. J. Agron. 18, 267–288 (2003). doi.org/10.1016/S1161-0301(02)00108-9 followed with modular architecture suited to diverse farming systems, while STICS7Brisson, N. et al. An overview of the crop model STICS. Eur. J. Agron. 18, 309–332 (2003). doi.org/10.1016/S1161-0301(02)00110-7 extended the framework to explicit organ-level carbon–nitrogen partitioning.
These systems now form the core infrastructure for global yield gap analyses, climate adaptation studies, and deployment risk quantification — increasingly coupled with remote sensing assimilation, machine learning emulators, and multi-model ensemble uncertainty frameworks.
Model Architecture
Genotype-Aware Parameters in Biophysical Models
The G×E×M framework — formalized by van Ittersum and Rabbinge5van Ittersum, M. K. & Rabbinge, R. Concepts in production ecology for analysis and quantification of agricultural input-output combinations. Field Crops Res. 52, 197–208 (1997). doi.org/10.1016/S0378-4290(97)00037-3 — recognizes that observed yield is the product of genetic potential, environmental expression, and management decisions. In biophysical models, genotype is encoded through cultivar coefficients: discrete parameters specifying photoperiod sensitivity, thermal time requirements for key phenological stages, maximum leaf area expansion, radiation-use efficiency, and grain-filling duration.1Jones, J. W. et al. The DSSAT cropping system model. Eur. J. Agron. 18, 235–265 (2003). doi.org/10.1016/S1161-0301(02)00107-7
Hammer et al.8Hammer, G. L. et al. Models for navigating biological complexity in breeding improved crop plants. Trends Plant Sci. 11, 587–593 (2006). doi.org/10.1016/j.tplants.2006.10.006 demonstrated that cultivar-parameterized models can navigate biological complexity across breeding populations and predict performance in unobserved environments — a capacity that purely statistical approaches cannot replicate without far larger empirical datasets. This extrapolation potential underlies Aluvial's spatial deployment intelligence: coefficients estimated from controlled calibration trials are propagated through high-resolution weather and soil surfaces to generate deployment risk profiles across thousands of geographies.
Advances in genomic prediction have further linked quantitative trait loci (QTL) to model parameters, enabling hybrid approaches that use marker data to inform priors on biophysical coefficients.9Cooper, M. et al. Predicting the future of plant breeding: complementing empirical evaluation with genetic prediction. Crop Pasture Sci. 65, 311–336 (2014). doi.org/10.1071/CP1400710van Eeuwijk, F. A. et al. Modelling strategies for assessing and increasing the effectiveness of new phenotyping techniques in plant breeding. Plant Sci. 282, 23–39 (2019). doi.org/10.1016/j.plantsci.2018.06.018 Genotype-aware models now provide the analytical substrate for connecting breeding pipeline outputs directly to commercial deployment risk.
Evidence Framing & Crop Model Library
Evidence Framing
Before running a single simulation, the evidence-framing stage defines the exact commercial decision and the evidence required to support it. Rather than conducting open-ended research, Aluvial synthesizes existing trial, field, and environmental data to establish what is known — and what is not — about a product's performance under real deployment conditions. Each product is mapped into a structured Genotype × Environment × Management (G×E×M) coordinate system,5van Ittersum, M. K. & Rabbinge, R. Concepts in production ecology for analysis and quantification of agricultural input-output combinations. Field Crops Res. 52, 197–208 (1997). doi.org/10.1016/S0378-4290(97)00037-38Hammer, G. L. et al. Models for navigating biological complexity in breeding improved crop plants. Trends Plant Sci. 11, 587–593 (2006). doi.org/10.1016/j.tplants.2006.10.006 identifying criteria for ideal responder classes and flagging critical knowledge gaps. This targeted knowledge base directly informs Aluvial's AI pipelines and baseline modeling.
Biophysical Prior
Once framed, this knowledge base feeds Aluvial's BMST — high-performance biophysical crop modeling software that we author and maintain — configured for major commodity grains, oilseeds, and cover crops. This mechanistic representation serves as a biophysical prior17Willard, J. et al. Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Comput. Surv. 55, 1–37 (2022). doi.org/10.1145/3514228 that enforces known physiological limits before any statistical learning begins, ensuring that deep learning layers are calibrated against residual behavior rather than inferring basic biology from scratch.16Feng, P. et al. Dynamic wheat yield forecasts are improved by a hybrid approach using a biophysical model and machine learning technique. Agric. For. Meteorol. 285–286, 107922 (2020). doi.org/10.1016/j.agrformet.2020.107922
Data Isolation
Client engagement data is never merged into shared calibration datasets or used to train global models for other accounts. Genotype identity tracking is maintained for all client-supplied materials, and each engagement remains logically isolated. Proprietary observations are returned exclusively within client-specific outputs.
Biophysical Models
The Mechanistic Engine: Physics-Informed Neural Networks
Aluvial's biophysical model library is configured for major row crops and bioenergy feedstocks — corn, soybeans, spring wheat, canola, and camelina. Daily biomass accumulation is simulated using a radiation use efficiency (RUE) framework11Monteith, J. L. Climate and efficiency of crop production in Britain. Philos. Trans. R. Soc. Lond. B 281, 277–294 (1977). doi.org/10.1098/rstb.1977.014012Sinclair, T. R. & Muchow, R. C. Radiation use efficiency. Adv. Agron. 65, 215–265 (1999). doi.org/10.1016/S0065-2113(08)60914-1 — where intercepted photosynthetically active radiation is converted to dry matter at a species-specific efficiency coefficient — while phenological progression is driven by growing degree day (GDD) thermal accumulation.13McMaster, G. S. & Wilhelm, W. W. Growing degree-days: one equation, two interpretations. Agric. For. Meteorol. 87, 291–300 (1997). doi.org/10.1016/S0168-1923(97)00027-0
Biomass production is constrained at each daily timestep by independent, multiplicative physiological stress scalars governing:
- water availability
- temperature stress
- nitrogen limitation
- phosphorus limitation
- potassium limitation
Each scalar operates on a strict index between zero and one. This multiplicative formulation ensures that a single severe constraint — such as a flash drought — can realistically suppress the crop's response to otherwise favorable conditions, mirroring yield losses observed under combined water and nutrient stress.14Steduto, P., Hsiao, T. C., Fereres, E. & Raes, D. Crop Yield Response to Water. FAO Irrigation and Drainage Paper No. 66. FAO, Rome (2012). fao.org/3/i2800e/i2800e.pdf Temperature functions also govern both the instantaneous photosynthetic rate and cumulative development, enforcing threshold-based reductions during extreme heat or cold.15Asseng, S. et al. Rising temperatures reduce global wheat production. Nat. Clim. Change 5, 143–147 (2015). doi.org/10.1038/nclimate2470
PINN Architecture
Bridging Biophysics and Deep Learning
The physical equations constrain the hypothesis space to biologically possible outcomes before any field data is introduced, while neural network layers are calibrated against the residual behavior — the complex, non-linear G×E×M interactions5van Ittersum, M. K. & Rabbinge, R. Concepts in production ecology for analysis and quantification of agricultural input-output combinations. Field Crops Res. 52, 197–208 (1997). doi.org/10.1016/S0378-4290(97)00037-38Hammer, G. L. et al. Models for navigating biological complexity in breeding improved crop plants. Trends Plant Sci. 11, 587–593 (2006). doi.org/10.1016/j.tplants.2006.10.006 that the mechanistic system cannot fully explain on its own. This is a Physics-Informed Neural Network (PINN) framework:17Willard, J. et al. Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Comput. Surv. 55, 1–37 (2022). doi.org/10.1145/3514228 rigid biophysics provides structure, deep learning provides empirical adaptation.
Unconstrained machine learning must simultaneously infer biological structure and environmental response from noisy field observations — requiring datasets that are often unattainable at early product stages. The physics-informed prior eliminates this requirement, compressing the data volume needed for precision by up to 2–3×17Willard, J. et al. Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Comput. Surv. 55, 1–37 (2022). doi.org/10.1145/3514228 and drastically improving extrapolation to novel geographies and management regimes where purely statistical systems degrade.16Feng, P. et al. Dynamic wheat yield forecasts are improved by a hybrid approach using a biophysical model and machine learning technique. Agric. For. Meteorol. 285–286, 107922 (2020). doi.org/10.1016/j.agrformet.2020.107922
Micro: Physiological Modeling at the Process Level
Field trials separate treatment effects from environmental noise slowly — over multiple seasons, at high cost.21Piepho, H.-P., Möhring, J., Melchinger, A. E. & Büchse, A. BLUP for phenotypic selection in plant breeding and variety testing. Euphytica 161, 209–228 (2008). doi.org/10.1007/s10681-007-9449-8 Aluvial's Micro infrastructure shortens that path by resolving crop physiology at the process level, inside controlled environments, before field budgets are committed.
Architecture — Physics-Informed Neural Networks
Aluvial uses Physics-Informed Neural Networks (PINNs) to model discrete physiological processes as a connected chain — from root water uptake through xylem transport, mesophyll conductance, and stomatal response.17Willard, J. et al. Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Comput. Surv. 55, 1–37 (2022). doi.org/10.1145/3514228 Each process layer is constrained by known biophysical relationships, so the model cannot fit data in ways that violate plant physiology. Hidden states — internal variables like leaf water potential and ABA signaling intermediates that are not directly observable — are inferred from the observed data rather than assumed away.
This architecture means the treatment effect of a biological input or genotype can be isolated at the process it actually acts on, not averaged across a whole-plant or whole-season response.
High-Resolution Data Capture
Greenhouse and growth chamber programs generate 3–4 replicated calibration cycles annually. Sub-daily environmental covariates — temperature profiles, vapor pressure deficit (VPD), photosynthetically active radiation (PAR), and soil moisture dynamics — are captured at the resolution the physiological models require.22Poorter, H. et al. Pot size matters: a meta-analysis of the effects of rooting volume on plant growth. Funct. Plant Biol. 39, 839–850 (2012). doi.org/10.1071/FP12049 Controlled nitrogen gradients and simulated stress events, including flash droughts and heat spikes, establish the environmental forcing conditions that define each calibration sprint.
Disentangling Signal from Noise
Trials are executed against immutable genotype panels spanning core reference lines, client germplasm, and exploratory material. Because the physiological model constrains how environment propagates through the plant system, the biological treatment effect can be isolated from random weather realizations and environmental covariance — not statistically averaged away, but mechanistically separated.23Skirycz, A. & Inzé, D. More from less: plant growth under limited water. Curr. Opin. Biotechnol. 21, 197–203 (2010). doi.org/10.1016/j.copbio.2010.03.002
Accelerating the Go / No-Go Decision
Calibrated response surfaces begin forming within a single growing season. The output is decision-grade intelligence: a high-value Fast No that halts wasted field spend on non-transferable signals, or explicit responder classifications that de-risk product launches before broad field budgets are committed.
A high-value Fast No stops wasted field spend on non-transferable signals before broad deployment budgets are committed.
Aluvial Macro: Field-Scale Yield Response via Hierarchical Modeling
Micro calibration establishes how a product or genotype behaves under controlled physiological conditions. Macro translates that into actionable spatial intelligence — where the product works, across real production geographies, at the resolution commercial decisions require.
Architecture — Physics-Informed Hierarchical Modeling (PIHM)
The Macro engine is built on a Physics-Informed Hierarchical Modeling (PIHM) framework17Willard, J. et al. Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Comput. Surv. 55, 1–37 (2022). doi.org/10.1145/3514228 that systematically bridges controlled R&D data and heterogeneous field variability. At its base sits the mechanistic biophysical prior from BMST. A Bayesian hierarchy pools observations across environments to estimate genotype- and management-conditional residual behavior, scaling evidence from controlled environments to commercial geographies through three progressively richer levels.19Malosetti, M., Ribaut, J.-M. & van Eeuwijk, F. A. The statistical analysis of multi-environment data: modelling genotype-by-environment interaction and its genetic basis. Front. Physiol. 4, 44 (2013). doi.org/10.3389/fphys.2013.00044
Level 1 — Foundation Weights
Establishes the baseline environmental fingerprint using large-scale climate and soil context. Encodes the prior expectation of crop behavior before any client data is introduced.3Muñoz-Sabater, J. et al. ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data 13, 4349–4383 (2021). doi.org/10.5194/essd-13-4349-2021
Level 2 — Micro-Calibrated Posteriors
Integrates controlled greenhouse data from Aluvial Micro to establish genotype-conditional response functions and treatment effects under defined stress conditions — updating the posterior with high-signal, low-noise observations.19Malosetti, M., Ribaut, J.-M. & van Eeuwijk, F. A. The statistical analysis of multi-environment data: modelling genotype-by-environment interaction and its genetic basis. Front. Physiol. 4, 44 (2013). doi.org/10.3389/fphys.2013.00044
Level 3 — Macro Field Observations
Pools field-level observations and in-season weather realizations to update the prior structure, capturing how treatments perform across complex, real-world Genotype × Environment × Management (G×E×M) interactions5van Ittersum, M. K. & Rabbinge, R. Concepts in production ecology for analysis and quantification of agricultural input-output combinations. Field Crops Res. 52, 197–208 (1997). doi.org/10.1016/S0378-4290(97)00037-3 at commercial scale.
High-Resolution Spatial Intelligence
Spatial outputs are generated at multiple resolutions tailored to commercial deployment scope: approximately 4 km continental US coverage using PRISM climate surfaces,18Daly, C. et al. Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. Int. J. Climatol. 28, 2031–2064 (2008). doi.org/10.1002/joc.1688 field-scale resolution using gSSURGO soil constraints, and approximately 9 km global analyses using ERA5-Land.3Muñoz-Sabater, J. et al. ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data 13, 4349–4383 (2021). doi.org/10.5194/essd-13-4349-2021 This allows the Macro engine to map responder terrains and identify geographies where yield, oil content, and profitability thresholds clear together.
Probabilistic Risk Outputs (P10 / P50 / P90)
Point estimates are insufficient for commercial risk decisions. Every Macro spatial output is delivered as a probabilistic surface20Asseng, S. et al. Uncertainty in simulating wheat yields under climate change. Nat. Clim. Change 3, 827–832 (2013). doi.org/10.1038/nclimate1916 — quantifying yield response, carbon intensity, and downside risk as explicit P10/P50/P90 confidence distributions. Procurement and commercialization teams get a quantified view of worst-case exposure before capital or field budgets are committed.
The Compounding Data Advantage
Each additional calibration cycle tightens uncertainty bounds for the target geography — physically sharpening parameter estimates across the entire crop system.19Malosetti, M., Ribaut, J.-M. & van Eeuwijk, F. A. The statistical analysis of multi-environment data: modelling genotype-by-environment interaction and its genetic basis. Front. Physiol. 4, 44 (2013). doi.org/10.3389/fphys.2013.00044 A third deployment cycle within a region produces materially narrower confidence intervals than an initial engagement over the same landscape. That is measurable uncertainty reduction driven by accumulated evidence, not cosmetic smoothing.
Carbon Intensity Translation: Carbon Prior + CI Engine
Carbon Prior Architecture
Biophysical outputs from the Macro engine do not produce a single CI score. They feed a causal model — Aluvial's Carbon Prior — that represents the full pathway from field management decisions to grams of CO₂-equivalent per megajoule (gCO₂e/MJ) as a directed acyclic graph (DAG).
Each node in the DAG corresponds to a discrete causal input: nitrogen application rate, yield response, oil content, residue management, soil carbon dynamics, and fuel conversion efficiency. Each carries independently attributed uncertainty. The result is a CI distribution, not a point estimate — and the distribution reflects the actual sources of variance in the pathway, not a sensitivity analysis applied after the fact.
CI-at-Risk (CIaR)
Beyond expected CI, the system computes CI-at-Risk (CIaR) — a downside-risk metric quantifying worst-case CI exposure under adverse environmental realizations. P10 conditions, including drought stress, heat events, and low-yield years, propagate through the full causal pathway, producing a stress-year CI distribution alongside the median. For procurement and compliance teams, CIaR answers the question that expected CI cannot: does this supply shed still qualify in a bad year?
The CI Engine
The CI Engine queries the Carbon Prior at runtime against the regulatory framework relevant to the decision. GREET, CA-GREET, CORSIA, and EU RED III are parameterized dynamically — so the same simulation outputs support feedstock sourcing, SAF qualification, and international compliance workflows without model re-execution. Switching from a 45Z evaluation to a CORSIA screen does not require a new model run. It requires a different query.
Additionality and 45Z Eligibility
Additionality is derived through counterfactual comparison against regulatory baselines — enabling direct estimation of eligibility under 45Z (IRA §13204) and EU RED III additionality provisions. The baseline is modeled from the same biophysical priors used for the treatment pathway, ensuring the delta reflects agronomic reality rather than a static regulatory default.
MRV Integration
For recurring compliance workflows, Aluvial integrates Sentinel-1 and Sentinel-2 anomaly detection to validate that claimed management practices occurred spatially and temporally as reported. This closes the loop between modeled CI claims and field-level verification — converting point-in-time assessments into defensible, auditable MRV infrastructure.
CI & Feedstock Intelligence: From Supply Shed to Compliance Claim
For feedstock developers and refiners, the relevant question is rarely “what is the expected CI score?” Expected CI tells you where a supply shed looks viable on paper. The operational question is what happens to that viability under adverse conditions — and whether a contracted supply shed qualifies under the framework that governs the fuel pathway being underwritten.
The Downside Tail Is the Contract Risk
In most procurement environments, the P10 CI outcome under drought-year conditions is more consequential than the median estimate. A supply shed that clears 45Z at P50 but fails at P10 is a compliance liability in any stress year. Aluvial's county-level CI screen evaluates supply sheds against the full probability distribution — identifying geographies where the program qualifies at P10, not just on average, before acreage or procurement contracts are committed.
Screening Feedstock Geographies Before Capital Moves
CI & Feedstock Intelligence maps where yield, oil content, carbon intensity, and profitability clear together — county by county, framework by framework. A camelina program targeting 45Z qualification in the northern plains looks different under CORSIA or EU RED III. A supply shed viable for one framework may carry disqualifying CI exposure under another. The CI Engine resolves these simultaneously, so procurement teams can identify geographies that qualify across the frameworks relevant to their fuel pathway before contracts lock in supply sheds that may not clear.
Long-Term Contract Viability Under Stress Scenarios
Long-duration feedstock contracts commit capital against yield and CI assumptions that may not hold across weather realizations. Aluvial models contract viability by geography across stress-year scenarios — surfacing the geographies where yield and CI remain above threshold in P10 conditions, and the geographies where a single adverse season breaks the program economics. That is the analysis that converts a supply shed map into a defensible procurement strategy.
What you receive
County-level P10/P50/P90 CI distributions under GREET, CA-GREET, CORSIA, and EU RED III. Yield and oil content surfaces by geography and stress scenario. Additionality estimates against regulatory baselines. Contract viability screens by supply shed and weather realization. MRV subscription tiers for ongoing compliance infrastructure.
Explore the Data
Sampled records from Aluvial's biophysical models showing where yield signals are present across commodity and oilseed systems.
Every major crop, every field, every year — all consistently modeled
Sampled records from Aluvial’s biophysical models showing where yield signals are present across commodity and oilseed systems.
Data That Compounds Across Programs
Every engagement contributes to tighter G×M response surfaces across the system. The mechanism is specific:
Over time, this creates a compounding intelligence effect: every validated observation improves the precision of subsequent analyses across connected deployment pathways.
What a Technical Validation Engagement Looks Like
The methodology above describes how the models are built. A first engagement puts them to work on a named commercial decision — scoped to the geographies, crop systems, and success criteria that matter to your program.
A typical first engagement covers 1–3 crop systems across 5–8 target geographies, delivering deployment suitability analysis, uncertainty characterization, and go/no-go placement recommendations against the decision you bring in.