PaperlayerView exact revision on arXiv

Source-linked AI summary

Agri-SCHOLAR: Simulation-Grounded In-Season Rolling Advisories for Agricultural Crops Using LLM-Bounded Optimization

Geetha Charan, Vedant Balasubramaniam, Manojkumar Patil, V. Priyanka, Kodur Sai Vinay Sathvik, Rohit P Suresh, Mohammad Sufyan Azam, Y. Narahari

arXiv:2610.31003v1cs.AIcs.MA

TL;DR

Existing agricultural advisories are often expert-driven and regional, while simulator-validated season-long plans require weather information unavailable in advance. Agri-SCHOLAR combines rolling weather updates with LLM-bounded optimization evaluated in APSIM; across 11 South Asian maize seasons, it beat the regional package of practices in every season.

  • Problem

    Existing advisories are expert-driven and regional rather than farm-specific, while simulator validation of season-long plans requires full-season weather that is unavailable for prospective use.

    Source: §1 · pp. 1–2

  • Method

    Agri-SCHOLAR rebuilds a weather trajectory from observations, short-range forecasts, and analog-year extrapolation, then uses LLM-proposed feasible action bounds and APSIM-guided optimization for each rolling advisory.

    Source: §1 · p. 2, §5 · p. 12

  • Results

    Across 11 maize seasons, Agri-SCHOLAR beat the regional package of practices in every season; mean yield rose by 15.2% in the irrigated region and 22.4% in the rainfed region.

    Source: §5 · p. 12

  • Takeaways & Limitations

    The results support rolling, simulator-grounded advisories as a way to address limitations of existing advisory systems without an accurate season-long weather forecast.

    Source: §1 · p. 2

    The simulator and reward omit field constraints such as minimum evenly spread dose and crop-stage suitability, so optimizer reliability depends on what they capture.

    Source: §3.2 · p. 20

Abstract

from arXiv · show

Existing agricultural advisories to farmers suffer from several limitations: (a) remain largely expert-driven, (b) do not adequately account for in-season weather, and (c) do not scale to individual farms. In this paper, our objective is to close these gaps by developing Agri-SCHOLAR (Simulation-Grounded Crop Harvest Optimization with LLM-Guided In-Season Advisory Rollouts). Agri-SCHOLAR is a rolling, in-season advisory framework that issues, at intervals of every few days or every week, a personalized advisory using the following information: soil and crop properties, observed weather to date, short-range forecast of weather, and a season extrapolation of weather based on historical analogs. The novelty and depth of Agri-SCHOLAR lies in each advisory being produced by a closed-loop system which couples multiple LLM agents with a powerful crop simulator (such as APSIM in our case). The simulator crucially serves as an evaluation oracle. In Agri-SCHOLAR, the LLM agents propose agronomically sound and feasible bounds on actionable parameters and an optimizer searches within those bounds against a multi-objective reward spanning yield, resource utilization, and economic feasibility. We conduct a detailed 11-year analysis of the maize crop in two selected locations of South Asia and show that Agri-SCHOLAR consistently outperforms the region's standard packages of practice and satisfactorily addresses the limitations of existing agricultural advisories.

1 Introduction

Agri-SCHOLAR addresses the mismatch between broad, expert-driven regional advisories and farm-specific needs by combining rolling weather updates with simulator-grounded optimization. In 11 years of South Asian maize evaluation, it consistently outperformed the regional package of practices.

1 Introduction

  • Existing advisories are often expert-driven, regional rather than farm-specific, and not checked against mechanistic crop response; farms within a zone differ in soil, cultivar, sowing date, and management.Earlier simulator-validated season-long plans require full-season weather in advance, which forecasting systems cannot provide at the fidelity needed for prospective deployment.Source: §1 · pp. 1–2
  • Each advisory is recalculated every few days to a week using observed weather, a short-range forecast, and an analog-based extrapolation for the rest of the season.New observations update the next advisory, so errors from earlier extrapolated weather do not accumulate over the season.Source: §1 · p. 2
  • LLM agents set agronomically feasible bounds on irrigation and fertilizer actions, while an optimizer searches within them using APSIM as an evaluation oracle.The optimizer evaluates candidates against a multi-objective reward covering yield, nitrogen and water-use efficiency, and economic feasibility.Source: §1 · p. 2
  • Agri-SCHOLAR consistently outperformed the regional package of practices across 11 years of South Asian maize evaluation.The comparison indicates the framework can issue rolling advisories without an accurate season-long weather forecast.Source: §1 · p. 2
  • The empirical demonstration covers 11 maize seasons across two contrasting irrigated and rainfed South Asian environments and reports improved economic and environmental outcomes alongside yield gains.An ablation study examines the contribution of agronomic LLM bounds.Source: §1 · p. 2

2 Related Work

Prior agricultural LLM systems primarily ground responses in retrieved text, while simulator-grounded approaches and simulator-based policy learning address different parts of the problem. Agri-SCHOLAR combines crop simulation, feasible bounded search, rolling weather updates, and farm-level advisories.

2 Related Work

  • Retrieval-grounded agricultural LLMs and their QA benchmarks assess language or multimodal performance, but do not check whether recommendations are physiologically viable for crops.Examples include ShizishanGPT [2], AgriGPT [3], AgriRegion [4], AgroAskAI [5], and benchmarks AGXQA, AgriBench [7], AgMMU, and AgEval.Source: §2 · pp. 2–3
  • Simulator-grounded predecessors include DSSAT-LM and Agri-SAGE [1], with Agri-SAGE comparing reasoning strategies over a 10-year retrospective in one region.Agri-SAGE couples multi-agent reasoning with APSIM.Source: §2 · p. 3
  • Agri-SCHOLAR differs from Agri-SAGE by making weekly decisions with information available at the time, bounding rather than directly proposing actions, optimizing profit rather than yield alone, and evaluating irrigated and rainfed settings.Agri-SAGE instead uses whole-season observed weather for a season-long plan and evaluates one region.Source: §2 · p. 3
  • Simulator-based learning systems train management policies but output numeric schedules rather than explainable, farmer-facing advisories.Work also reports that policies trained under normal conditions must be retrained for extreme weather [15].Source: §2 · p. 3
  • Operational services issue short-horizon, crop-stage guidance as weather evolves, but remain expert-driven and administrative-region based.The cited examples include India, China, and Vietnam.Source: §2 · p. 3
  • Agri-SCHOLAR combines simulator grounding with analog-year weather extrapolation, LLM-proposed feasible bounds, multi-objective scoring, rolling advisories, and farm-level personalization.The approach adds mechanistic validation and farm-level personalization to operational services that issue regional advisories.Source: §2 · p. 3

3 Methodology

Agri-SCHOLAR uses five agents to set context and constraints, search for weekly management actions, and verify recommendations through APSIM. Its optimizer maximizes cumulative simulated profit while the ABA’s agronomic bounds constrain the search.

3 Methodology

  • The five-agent pipeline combines preseason context-setting with weekly advisory production, then issues and logs each verified recommendation for later cycles.The PSA operates at season start; the other four agents jointly produce weekly advice.Source: §3 · p. 3

3.1 Context Retrieval Agent (CRA)

  • The CRA retrieves agronomic literature and regional guidance using a query formed from numerical weather, soil, and crop-stage features.It supplies downstream agents with a compact natural-language context bundle.Source: §3.1 · p. 3

3.2 PreSeason Agent (PSA)

  • The PSA sets the week-0 plan, including sowing date, mulch, and basal DAP, with sowing triggered by a three-day rainfall total of at least 20 mm.It selects the first qualifying window from June 1 to July 15 and uses its last day as the sowing date.Source: §3.2 · p. 4

3.3 Advisory Bounding Agent (ABA)

  • The ABA proposes agronomically plausible bounds that shrink the OA’s search space and prevent implausible actions from exploiting APSIM.Bounds are grounded in literature before the simulator is queried.Source: §3.3 · p. 4

3.4 Optimization Agent (OA)

  • The OA maximizes cumulative APSIM-simulated profit by optimizing weekly fertilizer and irrigation actions within ABA bounds.Previously issued actions are fixed; only the current week’s action is committed, while tentative later schedules are re-optimized at the next weekly decision point.Source: §3.4 · p. 4

3.4.1 Profit-Based Objective

  • The objective subtracts production costs and sustainability penalties from grain revenue, with APSIM supplying productivity and environmental indicators.The penalty combines in-season waterlogging and nitrogen oversupply with post-season drainage, leaching, and soil-carbon impacts.Source: §3.4.1 · pp. 5–6

3.5 Verification and Feedback Agent (VFA)

The VFA evaluates each candidate action in APSIM and returns crop and environmental metrics to support reward calculation and bound revision.

3.5 Verification and Feedback Agent (VFA)

  • The VFA returns simulated yield, soil-water status, nitrogen balance, drainage, and leaching to the OA for reward calculation.It runs APSIM on the weather file constructed for the current cycle.Source: §3.5 · p. 6
  • The VFA also returns simulation metrics to the ABA, which can tighten or relax bounds after repeatedly infeasible or dominated candidates.This links simulator feedback to revisions of the feasible search region.Source: §3.5 · p. 6
  • Each candidate action is translated into APSIM tool calls and evaluated using the current cycle’s constructed weather file.The resulting metrics include both crop outcomes and environmental indicators.Source: §3.5 · p. 6

3.6 Iterative Loops

Agri-SCHOLAR nests candidate search, bound revision, and weekly reruns into three loops. Inner and middle loops refine the current search, while the outer loop refreshes the advisory with new weather information.

3.6 Iterative Loops

  • The inner OA–VFA loop proposes a candidate, scores it in APSIM, and updates the OA’s search.Source: §3.6 · p. 6
  • When inner-loop convergence or an unsatisfactory candidate prompts revision, the ABA uses VFA feedback to adjust bounds and restart the search.Source: §3.6 · p. 6
  • Each week, the full pipeline reruns from scratch with observed weather, a new short-range forecast, and an updated analog tail to produce a new advisory.Source: §3.6 · p. 6

3.7 Weather Array Construction

The APSIM weather trajectory combines observations through the decision date, a synthetic seven-day forecast, and a historical analog for the remaining season. The forecast adds lead-time-growing stochastic errors, while analog selection uses only historical years and current-season information available by that date.

3.7 Weather Array Construction

  • The weather file joins pre-sowing and current-season observations, a perturbed seven-day forecast, and an analog-year tail for the rest of the season.All segments other than the forecast use information available on the decision date, preventing seasonal information beyond the seven-day horizon from entering the trajectory.Source: §3.7 · p. 6

3.7.1 Synthetic Seven Day Forecast

  • Forecast errors grow with lead time; precipitation is simulated first, and precipitation-state changes condition temperature and radiation errors.The precipitation parameterization preserves expected precipitation while allowing relative uncertainty to grow with lead time; the radiation shift adjusts for forecast wet versus dry days.Source: §3.7.1 · p. 7

3.7.2 Analog Tail Selection

  • The analog tail uses the historically most similar season selected from years before the target year, based on standardized, weighted climate descriptors.Descriptors include ONI, cumulative precipitation, and maximum-temperature anomaly; excluding the target and later years prevents future weather records from entering the analog.Source: §3.7.2 · pp. 7–8

4 Experimental Results

Across 11 seasons in two South Asian maize regions, Agri-SCHOLAR improves yield and profit over the regional package of practices (PoP), while reducing irrigated-region drainage and leaching. The experiments also show how rolling recommendations respond to changing conditions and how agronomic bounds affect the optimizer’s recommendations.

4.1 Experimental Setup

  • The study simulates maize across 11 kharif seasons in irrigated and rainfed regions, comparing weekly decisions with a fixed PoP plan under the same actual seasonal weather.Advisories use only weather available at each decision date; the issued plan is replayed under true weather for evaluation.Source: §4.1 · p. 8, §4.1.2 · p. 8, §4.1.3 · p. 8

4.2 Yield

  • Mean yield rose 15.18% in the irrigated region and 22.43% in the rainfed region, while mean net profit rose 17.2% and 22.8%, respectively, over PoP.Agri-SCHOLAR outperformed PoP on yield in every season in both regions; net profit was higher in every season in both regions.Source: §4.2 · p. 9, Figure 2 · p. 9, §4.3 · p. 9
  • Yield gains varied with rainfall distribution: the irrigated-region gain reached 29.45% in dry-spell 2023 but was 6.57% in well-distributed-rainfall 2018.In 2023, Agri-SCHOLAR detected the dry spell and irrigated as soil moisture fell; PoP’s fixed calendar could not respond.Source: §4.2 · p. 9, Figure 2 · p. 9

4.3 Profit and Resource Use

  • Higher yield outweighed higher input costs: mean profit gains were $1,066 ha−1 in the irrigated region and $1,322 ha−1 in the rainfed region.The irrigated-region gain also included a $228 ha−1 lower sustainability penalty; the rainfed-region gain came almost entirely from higher yield.Source: §4.3 · pp. 9–10, Table 2 · p. 10

4.4 Sustainability

  • In the irrigated region, mean deep drainage fell 52.8% and mean NO3 leaching fell 52.7%; in the rainfed region, the mean penalty fell 2.8%.The rainfed reduction was limited, likely because rainfall-driven drainage without irrigation leaves less room for management improvement.Source: §4.4 · p. 10, Figure 4 · p. 10

4.5 Case Study

  • In 2023, Agri-SCHOLAR applied 180 kg ha−1 UreaN in three splits and one irrigation, yielding 11,028 kg ha−1 with 12.9 kg N ha−1 leached.PoP used five more irrigation events and about 77% more nitrogen, yet yielded 2,509 kg ha−1 less.Source: §4.5 · pp. 10–11, Figure 6 · p. 11

4.6 Ablation: Advisory Bounding Agent

  • Loose bounds raised mean yield 2.1% and profit 2.5% but increased the sustainability penalty 20.1%; without bounds, mean profit dropped 76%.Without bounds, the optimizer produced extreme input amounts; the Advisory Bounding Agent generates bounds from retrieved context to keep recommendations practical and avoid simulator exploitation.Source: §4.6 · p. 12, Table 3 · p. 12

5 Conclusion and Future Work

Agri-SCHOLAR produces farm-specific advisories by rebuilding its weather trajectory weekly and optimizing feasible irrigation and fertilizer actions in APSIM. Across 11 maize seasons in two South Asian regions, it outperformed the regional package of practices, while field validation and prospective farmer pilots remain future work.

5 Conclusion and Future Work

  • Weekly weather trajectories combine observations to date, a short-range forecast, and an analog-year extrapolation for the remainder of the season.The approach avoids requiring an accurate season-long forecast.Source: §5 · pp. 12–13
  • LLM agents bound feasible irrigation and fertilizer actions, which an optimizer searches using APSIM as its evaluation oracle.The reward is profit-based and includes sustainability penalties.Source: §5 · p. 12
  • Across 11 maize seasons in one irrigated and one rainfed South Asian region, Agri-SCHOLAR beat the regional package of practices in every season, even when that baseline used perfect knowledge of true weather.Mean yield rose 15.2% and 22.4%, and mean net profit rose 17.2% and 22.8%, in the irrigated and rainfed regions, respectively.Source: §5 · p. 12
  • Future work includes calibrating and validating the simulator against field trials, replacing emulated forecasts with operational weather products, and running prospective farmer pilots.The authors also identify sowing-date decisions and learning reward weights as directions for further study.Source: §5 · p. 13

1 Cost and Penalty Parameters

The optimization reward uses local-currency prices and penalties to represent production costs and monetized environmental losses. Price inputs reflect prevailing effective prices, while environmental parameters assign costs to specified resource losses and crop failure.

1 Cost and Penalty Parameters

  • The optimization agent uses fixed cost and penalty parameters expressed in local currency units, with dollar values converted post-experiment using purchasing power parity.The simulator uses local-currency values during experimentation.Source: §1 · p. 16, Table 4 · p. 16

1.1 Price Parameters

  • Prices reflect prevailing market, administered, or farmer-paid rates, including applicable subsidies and pricing policies.The grain support price is 20.90 LCU/kg, and the effective urea-N price is 13.04 LCU/kg of elemental N.Source: §1.1 · p. 16
  • The effective DAP price is 27.00 LCU/kg, labor costs 350 LCU per operation, and irrigation water costs 2.50 LCU/mm.Retained crop residue is valued at 2.00 LCU/kg.Source: §1.1 · p. 16

1.2 Environmental Parameters

  • Environmental penalties monetize waterlogging at 5,000 LCU/ha-season, deep drainage at 15.00 LCU/mm, and nitrogen leaching at 25.00 LCU/kg of leached N.These parameters represent partial waterlogging costs and the combined economic effects of deep drainage and associated nutrient losses.Source: §1.2 · p. 17
  • The model assigns a 2,000 LCU penalty to nitrogen applied beyond estimated crop requirements and 150 LCU/kg of carbon for soil organic carbon loss.The SOC-loss value is based on the social cost of carbon.Source: §1.2 · p. 17
  • Crop failure is assigned a cost of 50,000 LCU/ha, based on projected maize production costs and yield.The cited 2023–24 production cost and projected yield imply approximately 51,522 LCU/ha.Source: §1.2 · p. 17

2 Case Study: Full Weekly Advisories

The 2023 irrigated-region case study illustrates weekly advisories that adapt fertilizer and irrigation decisions to crop stage, rainfall, and soil moisture. The season used three urea applications and one irrigation, with the remaining water supplied by rain.

2 Case Study: Full Weekly Advisories

  • At sowing, the advisory prescribed 200 kg/ha DAP, 26 kg/ha residue mulch, and 128.3 kg/ha urea, with the urea applied as a side-dressing.The sowing date was 4 July 2023, when the monsoon had set in and soil was moist enough for germination.Source: §2 · p. 17
  • In three advisories, after 65 mm, 36 mm, and 101 mm of rain last week, fertilizer and irrigation were withheld because soil stayed moist and week-0 urea remained available; no water or N stress was expected.Source: §2 · p. 17
  • During later grain filling and maturation, the advisories generally withheld irrigation and fertilizer because soil moisture remained adequate or the crop was no longer expected to gain yield from added inputs.The week-8-to-10 advisories also withheld inputs while noting that recent rainfall and available week-7 urea remained relevant.Source: §2 · p. 18
  • The season summary reports 11,028 kg/ha yield from 391.4 kg/ha urea in three splits and one 63 mm irrigation event, with 12.9 kg N/ha nitrate leached.The one irrigation occurred in week 6; the remaining water came from rainfall.Source: §2 · pp. 18–19

3 Ablation of the Advisory Bounding Agent

The ablation compares the Advisory Bounding Agent (ABA) with loose or absent bounds in the irrigated region over 2015–2025. Removing bounds can produce attractive simulated metrics, but the unbounded optimizer also generates infeasible actions and severe environmental and profit losses.

3.1 Results

  • Loose Bounds’ nitrate leaching varied with weather: it reached 30.73 kg/ha in 2017, versus 5.80 kg/ha for Agri-SCHOLAR, and 22.52 versus 12.90 kg/ha in 2023.The authors report that heavy rain following a large early dose is associated with pronounced losses.Source: §3.1 · p. 19
  • Across seasons, Loose Bounds sometimes led simulated yield and profit, but its sustainability penalty was 20.1% higher than Agri-SCHOLAR’s, while No Bounds produced negative net profit in four seasons and the highest nitrate leaching in every season.The reported No Bounds leaching reached 694.18 kg/ha in 2025 and 3,270.66 kg/ha in 2019.Source: §3.1 · p. 19

3.2 Observations

  • Without ABA, the optimizer exploited gaps in APSIM and the reward, which did not represent minimum spreadable doses or whether timing and dose were agronomically appropriate.Fixed labor costs also encouraged schedules that looked efficient to the optimizer but were unsound on the farm.Source: §3.2 · p. 20
  • Examples included excessive nitrogen at planting and 2–4 kg/ha token top-ups that standard equipment could not spread evenly.In 2019, No Bounds applied 6,062.2 kg/ha UreaN at planting; a 2.8 kg/ha top-up was about 6% of a 45 kg bag.Source: §3.2 · p. 20
  • The 2019 No Bounds plan leached 3,270.66 kg/ha nitrate yet matched the bounded variants’ 6,927 kg/ha yield and lost $2,887/ha net.The authors caution that simulated gains do not establish better advice when the model and reward cannot assess agronomic quality.Source: §3.2 · p. 20
  • ABA restricted searches to doses and timings supported by agronomic guidance for the crop stage, yielding farmer-followable plans at the cost of small simulated-profit losses in some seasons.In the 2023 example, 180 kg/ha UreaN was split across sowing, rapid vegetative growth, and pre-tasselling.Source: §3.2 · p. 20
Loading 2610.31003v1…