A Proposed Architecture and Demonstration for Channel-Aware Calibrated Estimation of Yarn, Dye, and Chemical Requirements at Sarena Industries (Pvt.) Ltd.
We propose an AI-driven Material Requirements Planning architecture for Sarena Industries' vertically integrated woven-textile operations, designed to deliver calibrated estimates of yarn, dye, and chemical requirements for incoming orders across three of Sarena's four sales channels: global fashion-fabric export, workwear and protective-fabric export, and captive supply to the Sefam retail group. The fourth channel — full-package apparel from the Sarena Apparel SBU — operates on a methodologically distinct cut-and-sew BOM and is deferred to a subsequent engagement. The proposed architecture follows the seven-layer reference pattern observed across mature international mills (Arvind, Shenzhou, Toray, Milliken, Esquel), with each layer specifically adapted to Sarena's channel mix, product portfolio, and operational scale. We propose to demonstrate the architecture in two stages: an initial validation on synthetic data generated from published textile-engineering distributions and ASTM/ISO/AATCC specification standards, followed by calibration on Sarena's historical lot records once data access is established. Honest expected outcomes on calibration include 95% confidence-interval coverage on dye and chemical consumption within ±4 percent for in-family constructions, with explicit out-of-family flagging for novel constructions that defer to planner judgment. The document presents the channel scope, the architectural reference pattern, the practice survey of international mills, the proposed synthetic-data foundation, and the runtime path from order intake to requirement prediction. An accompanying interactive web demonstration provides a working illustration of the proposed system on representative synthetic data.
Keywords: material requirements planning; vertically integrated textile; calibrated estimation; yield coefficient learning; recipe management; dye-bath grouping; uncertainty quantification; channel-aware planning; Sarena Industries
Sarena Industries operates as a vertically integrated woven-textile producer with output flowing through four distinct sales channels. The channels share the same physical assets — weaving looms, range-dyeing machinery, finishing lines, chemical inventory — but differ materially in order characteristics, demand visibility, and planning cadence. We propose that any production-grade MRP system at Sarena must reason about all four channels simultaneously, because their constraints couple through the shared assets. Within the present engagement, however, we propose to scope the architecture to three of the four channels, deferring the fourth to a subsequent phase.
We propose to retain Channels A, B, and C for the present work because they share a common BOM structure — yarn, dyes, and process chemicals as the dominant inputs — and a common decision space at the wet-processing stage. A single architecture serves all three with channel-specific parameterization, rather than three siloed systems. Channel D requires a deeper BOM hierarchy and different optimization variables; we propose it as Phase 2 work.
The three in-scope channels carry meaningfully different MRP characteristics. The same physical asset base must service orders that differ in size, repeat structure, shade complexity, process depth, yield predictability, and demand-visibility horizon. The proposed architecture treats these differences as channel-aware parameters rather than separate codepaths.
| Characteristic | A: Fashion Export | B: Workwear & Protective | C: Captive Sefam |
|---|---|---|---|
| Typical order size | 5–50K metres | 2–20K metres | 10–100K metres (repeat) |
| Construction repeatability | Low to medium | Medium | High |
| Shade complexity | High (custom fashion) | Medium (catalogue & custom) | Medium (brand libraries) |
| Process complexity | Standard wet processing | High (multi-stage treatment) | Standard wet processing |
| Yield uncertainty | Medium | High | Low (repeat lots) |
| Demand visibility horizon | 4–12 weeks | 4–8 weeks | 12–26 weeks (rolling) |
| Dominant MRP failure cost | Lost margin on over-buy | Certification & delivery risk | Internal but reputational |
| Optimisation priority | Inventory turn, dye reuse | Process yield, compliance | Service level, throughput |
We propose that the system serve all three channels through a single planning core. Channel-aware behaviour is captured through configuration rather than code branching: per-channel demand-arrival processes, per-channel construction-family priors, per-channel optimisation objectives. This pattern is observed across all the international leaders surveyed in Section 4 and is the principal architectural decision that distinguishes a planning system from a planning tool.
Material Requirements Planning, at its operational core, answers a sequence of coupled questions for each incoming order and for the order portfolio as a whole. We propose to characterize the problem precisely before turning to the architecture, because the precision of the problem statement determines the defensibility of the solution.
For each incoming order, the system must answer four questions. First, what griege yarn is required, in what quantity, of what specification, to weave the requested fabric — accounting for warp and weft requirements, expected waste at weaving, and the historical variation in yield for the relevant construction family. Second, what dyes are required, in what quantity, to achieve the target shade on the target substrate — accounting for the shade's depth, the dye class, the substrate's affinity, the chosen machinery's liquor ratio, and the bulk-versus-lab dip divergence observed historically for similar shades. Third, what auxiliary chemicals are required across the process route — scouring agents, bleaching agents, wetting agents, salt, alkali, levelling agents, after-treatment chemicals, softeners, finish chemicals — at quantities dependent on substrate, recipe, and machinery. Fourth, with what uncertainty are each of these estimates being delivered, so that procurement decisions can be made with appropriate safety stock and so that out-of-confidence cases can be flagged for planner judgment rather than relied upon for automated decisions.
For the order portfolio as a whole, the system must additionally answer two further questions. Fifth, which orders can be grouped into shared dye baths to reduce dye consumption per metre, given the shade-compatibility and timing constraints. Sixth, what procurement actions are recommended this week, given the requirement estimates, current inventory positions, supplier lead times, and the cost of stocking out versus the cost of over-buying.
The framing of automated MRP as "exact calculation" is widely encountered but technically unsupportable. Shrinkage at finishing varies 2–8% on identical specifications run under nominally identical conditions; dye exhaustion diverges 5–15% between lab dip and bulk on familiar shades and more on novel ones; inspection cut-losses depend on operator decisions and humidity. The inexactness is physical, not informational, and no model removes it. We propose to deliver instead a system that is calibrated — its stated confidence intervals match realized outcomes — and auditable — every number is traceable to specific historical lots and named learned coefficients. On this standard the system is testable, defensible under audit, and substantially tighter than current manual estimates. This is the operational version of what was asked.
The mature international vertically integrated textile mills surveyed for this proposal converge on a seven-layer reference architecture for MRP. Specific implementations vary in the sophistication of each layer and in integration depth, but the layer structure itself is consistent across companies, vendors, and academic descriptions. We propose to follow this pattern at Sarena, adapting each layer to the channel mix and operational scale described in Section 1.
Standardized capture of construction (fibre composition, yarn count, weave structure, GSM, width), shade (Pantone or CIE Lab coordinates, lab-dip approval state), finish requirements, quantity, delivery date, and channel. Validation against feasibility rules at intake — does the requested construction map to an existing capability, is the shade within the chromatic range achievable on the requested substrate, is the lead time achievable. Integration with buyer-side PLM systems where present (Centric, Lectra, or buyer-custom). The layer's principal output is a fully specified order object with intake-time validation flags attached.
Conversion of the fabric specification into yarn requirements (warp count × density + weft count × density, with allowance for loom waste), dye requirements (substrate × shade × depth, mapped through a recipe library), and auxiliary chemical requirements (per process stage, parameterized by recipe and machinery). Each conversion applies a yield coefficient. The legacy industry approach treats yield coefficients as static engineering constants drawn from standards tables; the modern approach treats them as learned from historical lot data, with explicit uncertainty quantification per construction family and per process route.
Prediction of the actual consumption that will be observed in production, accounting for shrinkage, weight loss, recipe-bulk divergence, and process losses. The modern stack combines statistical models for yield variation (Beta or Gaussian distributions over historical observation), recipe-management systems for colour (target shade → predicted dye consumption from lab-dip plus a bulk-correction model fitted to historical bulk-versus-lab gaps), first-principles chemistry models for specific processes (especially valuable for novel constructions where statistical history is thin), and process digital twins for complex multi-stage operations. The layer's principal output is a posterior distribution over each material requirement, not a point estimate.
For repeat constructions and common yarn lots — relevant especially to Channel C and partly to Channel B — demand-driven MRP replaces traditional time-phased MRP. Buffer stocks of intermediate goods are sized based on demand variability and replenishment lead time, with explicit safety stocks calibrated to a chosen service level. Buffers are recalculated on a rolling basis as demand observation accumulates. The pattern is well-established in process industries and adapts cleanly to textile wet processing where many constructions repeat across seasons.
Once requirements are known with quantified uncertainty, the optimisation layer allocates yarn lots to orders, groups orders into shared dye baths (the combinatorially hard problem at the heart of dyehouse productivity), sequences machines, and schedules procurement to align with delivery dates. Mature mills use mixed-integer linear programming (Gurobi, CPLEX, or open-source CBC/HiGHS) or constraint programming (CP-SAT) for the principal optimisation. Some research-forward operations have shifted partial scheduling decisions to reinforcement learning, particularly for high-variance sequencing problems on bottleneck machinery. The layer's principal output is a recommended action plan with explicit costs and trade-offs for the planner to accept or modify.
Connection to the enterprise systems that own the data and execute the decisions. The textile industry's standard ERP backbone is SAP S/4HANA with the IS-Mill industry solution, used by Arvind and most large mills; alternative deployments include Microsoft Dynamics, Oracle Process Manufacturing, and custom builds. The MRP architecture sits as a recommendation engine alongside the ERP, reading order, inventory, and historical-lot data; writing recommended actions; and being explicitly subordinate to the ERP as the system of record. Integration also extends upstream to PLM and downstream to MES on the shop floor.
Actual material consumption from each lot feeds back into the yield-coefficient models, the recipe-management system, and the bulk-versus-lab correction model. Closed-loop learning where the system tightens its confidence intervals and improves its point estimates as new lot data accumulates. This is the layer where the recent advances in machine learning have had the most tangible impact in textile manufacturing — learning yield and consumption coefficients from data rather than relying on static engineering estimates published in standards tables. We propose this layer be designed in from the outset, even if its full power is only realized after several months of operational data.
We propose that Sarena's MRP system be built as one architecture with channel-aware parameterization, not three systems for three channels. The seven layers serve all three in-scope channels through configuration. Channel-specific behaviour — different demand-arrival processes, different priors over construction families, different optimisation objectives — is data, not code. This is the principal architectural decision that determines whether the system generalizes to future channels (including Channel D in Phase 2) or hardens into per-channel monoliths.
The mature mills surveyed below converge on the seven-layer architecture described in Section 3, differing principally in the sophistication of specific layers, the depth of integration with brand-customer systems, and the maturity of their continuous-learning loops. We summarize the practice of each below to anchor the proposed work in observable international practice rather than in unverified novelty. Specific implementation details inside each company are largely non-public; the descriptions below are accurate at the architectural-pattern level and conservative on implementation specifics.
Operates SAP S/4HANA with the IS-Mill industry solution as the ERP backbone. Has invested in machine learning for production planning and yarn forecasting since approximately 2018. Maintains internal tools for shade matching, recipe optimisation, and yield prediction. Integrates with multiple brand-customer planning systems as a Tier-1 supplier. The Human Protection Division (IGNX brand) operates specialized chemistry-aware recipe management for PROBAN, Pyrovatex, and aramid fabric systems — directly analogous to Sarena's PROBAN-licensed operations, with the methodology transferring substantially across the analog.
Probably the most operationally sophisticated vertically integrated mill in the world at its scale. End-to-end visibility from yarn intake to garment dispatch. Reportedly uses reinforcement learning for portions of the scheduling problem. Heavy investment in industrial IoT, sensor instrumentation, and process digital twins. Demand-driven planning with very tight integration to brand-customer demand signals. The differentiator is data-flow continuity and integration discipline rather than exotic algorithms — Shenzhou's competitive edge is operational rather than methodological.
Strong investment in first-principles chemistry modelling and recipe management. Less publicly visible on planning systems specifically. Substantial investment in fibre-to-finished-product traceability and process-control instrumentation. Likely the most advanced among the surveyed leaders on novel-material yield prediction, where pure machine-learning approaches struggle because of thin historical data. The first-principles approach complements statistical learning and is particularly relevant to Channel B at Sarena, where novel technical-fabric constructions appear regularly and lack the historical lot density that supports purely data-driven approaches.
Operates one of the world's largest internal textile research laboratories. Recipe management is a core organisational strength. Vertical integration across yarn to finished fabric enables tight process control and rapid iteration on novel recipes. Specific MRP architecture is not public, but the operating-excellence reputation suggests well-developed continuous-learning loops. Milliken's FR fabric business serves the same end markets as Sarena's PROBAN business and operates to the same international standards (NFPA 2112, EN 11611, EN 11612), making the Milliken benchmarks directly relevant to Channel B targets at Sarena.
Strong cotton-to-shirt traceability story. Machine-learning forecasting at multiple horizons. Sustainability metrics — water, energy, carbon per order — integrated into planning decisions. The sustainability-integrated MRP is a pattern especially worth studying for Sarena, whose ESG positioning (OEKO-TEX Made in Green, GOTS, ZDHC, Higg Index) would benefit substantially from the same integration. Tracking environmental impact at the per-order level is also a hard requirement for several of Sarena's European fashion customers.
Operates a structurally different MRP problem because solution-dyed acrylic production locks colour at the fibre extrusion stage, eliminating most downstream dye-related uncertainty. Studied here as a contrast case rather than a model Sarena would replicate. The contrast is instructive: Glen Raven's MRP is dramatically simpler than Sarena's specifically because colour is no longer a downstream variable. The lesson for Sarena is that the dyehouse — where colour is added late and where most yield and recipe uncertainty lives — is the structurally hardest layer and the layer where the most planning value can be unlocked.
The survey supports a clear conclusion. The architectural pattern is known. The methodological building blocks — yield-coefficient learning, recipe management, MILP-based dye-bath grouping, continuous-learning loops — are established. The differentiation among the international leaders sits in execution depth, integration discipline, and the quality of their continuous-learning feedback. We propose that Sarena's competitive position relative to peer mills is best advanced not by methodological novelty but by disciplined adoption of the established pattern, adapted to Sarena's channel mix, scale, and operational style.
We propose to develop, validate, and demonstrate the architecture on synthetic order and lot data generated from published textile-engineering distributions and international specification standards, with subsequent calibration on Sarena's actual historical lot records once data access is established. This staged approach is the standard practice in industrial-AI methodology development. It allows the architecture to be fully built and validated before Sarena's data integration is complete, and it produces a defensible methodological foundation that can be cited in any publication, regulatory review, or customer due-diligence enquiry. The synthetic data is for methodology development and demonstration; the production deployment will be calibrated on Sarena's real lot data.
We propose that the order, specification, and lot-data schemas be built directly against the international textile specification standards. The schemas inherit the field structure, tolerance definitions, and test methodology references from the standards, ensuring that any future integration with buyer PLMs, ERP systems, or certification bodies operates on a familiar data model. The principal standards we propose to reference:
We propose to model process variability using parameterized distributions drawn from the textile-engineering literature. Each distribution choice is named explicitly, supported by literature reference, and exposed as a configurable parameter so that calibration on Sarena's real data can subsequently update the parameters without changing the architecture.
| Variable | Proposed Distribution | Parameterization Driver | Literature Anchor |
|---|---|---|---|
| Shrinkage at finishing | Beta(α, β) | Construction family × finish route | Hu (2008); standard textile mechanics |
| Dye exhaustion | First-order kinetics + Gaussian noise | Dye class × substrate × machinery | Vickerstaff; Aspland |
| Lab-to-bulk dye divergence | Truncated Gaussian, μ ≈ 0, σ varies by class | Dye class × shade depth | AATCC technical bulletins |
| Scouring weight loss | Gaussian | Substrate composition | Textile chemistry handbooks |
| Weft insertion / pick density variation | Gaussian around target density | Loom type × yarn count | Lord & Mohamed; standard weaving |
| Inspection cut-loss | Weibull or log-normal | Quality grade × defect rate | ASTM D5430 inspection literature |
| Order arrival rate (Channel A) | Poisson with seasonal modulation | Fashion calendar × buyer mix | Apparel demand literature |
| Order arrival rate (Channel B) | Poisson, near-uniform | Industrial demand cycles | Workwear market reports |
| Order arrival rate (Channel C) | Quasi-deterministic rolling forecast | Retail-driven replenishment | Vertically integrated retail planning |
| Construction repeat pattern | Zipf-like over construction catalogue | Channel × season | Empirical textile-industry observation |
| Order size distribution | Pareto-like, channel-specific α | Channel × buyer tier | Empirical export-order distributions |
We propose the following generation pipeline, producing a labelled synthetic dataset that supports both architectural validation and demonstration. The dataset is constructed such that the ground-truth values for every yield coefficient, dye consumption, and chemical requirement are known, allowing rigorous evaluation of calibration coverage and prediction quality.
We propose to be explicit in all communication, internal and external, that synthetic data is for methodology development, architecture validation, and demonstration. Production deployment requires calibration on Sarena's actual historical lot records, which is the next phase of the engagement and the point at which the system becomes operationally meaningful. The synthetic-data phase establishes that the methodology is sound; the real-data phase establishes that it is calibrated for Sarena specifically.
We propose the following runtime workflow for each incoming order entering the system. The workflow operates on the trained yield-coefficient models, the recipe-management module, and the process simulator established during the synthetic-data phase, and subsequently re-trained on Sarena's real data during the calibration phase.
We propose that the system's principal validation metric be calibration coverage: the fraction of realized outcomes that fall within the stated confidence interval. A 95% confidence interval that captures 95% of realized values is properly calibrated; one that captures 75% is over-confident; one that captures 99% is under-confident and useless for procurement. Targeting calibration explicitly — rather than only point-estimate accuracy — is the standard the international leaders surveyed in Section 4 operate to, and is the standard we propose for the Sarena engagement.
We propose to demonstrate the runtime workflow on the synthetic dataset described in Section 5, with three core demonstrations:
An accompanying interactive web demonstration provides a working illustration of the proposed system on representative synthetic data, with channel selection, order entry, BOM explosion with uncertainty bands, recipe-management workflow, order-portfolio optimisation, and calibration metrics presented in an interactive interface. The web demonstration is the proposed artifact for the Lahore meeting.
We propose a three-phase engagement structure for the MRP work at Sarena.
Complete the architecture specification, implement the synthetic-data foundation, train the prediction modules on synthetic data, validate calibration coverage and out-of-family behaviour, deliver the interactive web demonstration, and present the demonstration to Sarena's operational and executive leadership. Deliverables: this proposal document, the working web demonstration, the architectural specification, and the validation report on synthetic data.
Establish data-access protocols with Sarena's PPC, ERP, and shop-floor systems. Extract a structured historical lot dataset covering yarn consumption, dye consumption, chemical consumption, realized yield, and recipe records. Re-train the prediction modules on Sarena's actual historical data. Re-validate calibration coverage on Sarena's own held-out lots. Adjust per-channel configurations to match observed Sarena patterns. Deliverables: Sarena-calibrated prediction modules, validation report against Sarena's historical lots, integration specification for Phase 3.
Deploy the system in shadow-mode alongside existing planning processes for a defined order subset (proposed: Channel C only initially, as the lowest-risk channel with the highest repeat structure). Operate the system in advisory mode where its recommendations are produced but procurement decisions remain manual. Establish the continuous-learning feedback loop. After demonstrated calibration in production over a defined evaluation window, expand to Channels A and B and to active recommendation mode. Deliverables: deployed system, operational metrics report, recommendations for Phase 4 expansion (Channel D, deeper integration, additional layers).
We propose to be explicit, with Sarena's leadership and within the engagement documentation, about what the proposed system will not do. It will not calculate "exact" material requirements, because exactness is physically unobservable and unsupportable in wet processing. It will not eliminate planner judgment, because out-of-family constructions and novel shades will remain — and should remain — escalation cases for human review. It will not replace the ERP, the PPC team, or domain knowledge built over decades at Sarena. The proposal is for a recommendation engine that delivers calibrated, auditable estimates with quantified uncertainty, integrated alongside the existing planning organisation, and improving steadily through observation. The architecture is established international practice, adapted to Sarena's specific channel mix, scale, and operational style.
Document version 1.0 · Prepared as a working draft for Sarena Industries leadership review · Accompanying interactive web demonstration provided as a separate deliverable · Subsequent versions will incorporate feedback from the Lahore review meeting and adjust scope accordingly