Table of Contents
A formulation is robust when it stays within specification across the normal spread of raw material lots and approved suppliers, not just with the lot used in development. You get there by measuring how incoming materials vary, testing the recipe against that variation on purpose, choosing the recipe where results move least, and adjusting dosages when a lot needs it.
Key takeaways
- Conforming raw materials can still produce a failing product. The FDA notes that this can happen even when materials meet pharmacopeial specifications, which generally address only chemical identity and purity.
- The effect can be large. In a Mississippi DOT study, a new sample of cement from the same source changed 28-day concrete strength by up to 21.6%, and a change of cement source by up to 31.8%.
- The recipe sets the sensitivity. In the same study, the percentage change was larger when 25% fly ash was in the mix.
- Treat lot properties as noise factors in experiments and as inputs in models, then choose the recipe that is least sensitive to them (NIST/SEMATECH).
- Single-property incoming limits misclassify lots when properties are correlated; multivariate acceptance regions built from plant data address this.
Why does a proven recipe fail with a new lot or supplier?
A supplier certifies what its specification lists. Your recipe depends on what it actually does in your mix, and the two are not the same thing.
The FDA's PAT guidance says that analytical methods for chemical attributes have progressed, but that some physical and mechanical attributes of ingredients are not well understood, so undetected variability of raw materials may show up in the final product. It adds that formulation strategies exist that are not affected by minor differences in physical attributes, but that they are often based on the experience of a particular formulator. That is exactly the knowledge that leaves with retiring staff.
Standard supplier tests do not always predict performance either. In the cement study, the authors found that trends in mortar cube strengths do not always reflect trends in concrete strengths. A material's own standard test may not predict how it behaves in your product.
How large can the effect be?
Published studies with real material lots show the scale.
| Raw material | Product | What was varied | What happened |
|---|---|---|---|
| Portland cement | Concrete (gravel aggregate) | 7 cements, 3 samples per source about one month apart, 84 mixes | 28-day strength changed by up to 21.6% between samples of one source and up to 31.8% between sources; changes grew with 25% fly ash (Varner, 2013) |
| Microcrystalline cellulose (MCC) | Granules for tablets | 12 MCC lots, including 6 made on purpose to vary more than normal production | Lot properties had an effect similar to or larger than the liquid-to-solid ratio, the most influential process setting (Portier et al., 2021) |
| Polymer resin | Industrial film | Resin properties from plant records | A PLS model related resin properties to film quality and was used to set a multivariate acceptance region (Duchesne and MacGregor, 2004) |
In the MCC study, the lot properties that mattered were water binding capacity, particle size distribution width, bulk density and moisture. They also affected processability: the granulator stops above a 20 Nm torque safety limit, so a difficult lot can stop the line, not just shift a result.
What does "robust" mean for a formulation?
ICH Q8(R2), the pharmaceutical development guideline, defines process robustness as the ability to tolerate variability of materials and changes of the process and equipment without a negative impact on quality. Its design space explicitly combines material attributes with process parameters.
Taguchi's robust design gives the practical framing. The NIST/SEMATECH e-Handbook describes the aim as making a product less variable against variation engineers cannot control, by separating design factors, which you choose, from noise factors, which you do not. For a formulation:
- Design factors: component dosages, order of addition, mixing time, temperature.
- Noise factors: lot properties of each raw material, supplier, storage age of a material.
A robust recipe is not always the one with the best average result. It is the one with the highest chance of meeting the full specification across the noise you will meet in production.
How do you measure how much your raw materials vary?
Start with records you already have, then fill the gaps.
- Collect certificates and incoming tests per lot. ICH Q7, the GMP guide for active ingredients, sets a useful discipline: full analyses on at least three batches before reducing in-house testing, periodic full analyses compared with the certificates, and regular checks of certificate reliability.
- Add the attributes that matter to your recipe. If performance depends on particle size or water absorption and the certificate reports only purity, you need your own test.
- Keep a retained sample of every lot, labelled with the lot number, so a problem can be traced back.
- Look at properties together. Lot properties are usually correlated. In the MCC study, three principal components captured 92% of the variability in the dataset.
Correlation is also why single-property limits fail. Paris, Duchesne and Poulin explain that univariate limits may lead to misclassification of lots. Loosening the limits to accept all past good lots raises the risk of accepting bad ones; tightening them rejects more good lots, which typically leads to higher purchasing costs.
How do you design a formulation that tolerates the variation?
The steps below combine robust design with the modelling approach described in our virtual lab guide.
- List the noise factors. For each raw material, note the supplier, the lot and the measured properties that could plausibly affect the result.
- Get lots that span the range. The MCC study used lots that the supplier prepared on purpose to vary more than standard production, then chose the most extreme lot in each quadrant of a PCA plot. Ask suppliers for lots near the edges of their specification, or keep retained samples of unusual lots.
- Cross recipe factors with lot factors. Taguchi's inner and outer arrays run a small lot design at each recipe setting; NIST's example of 8 recipe settings by 4 noise settings gives 32 runs. Compare recipes on both average and spread: the handbook notes that you cannot simply pick the winner.
- Block lots you are not studying. If one lot runs out halfway through a study, the results mix recipe and lot effects. NIST's rule applies: block what you can, randomize what you cannot.
- Put lot properties into the model. A model that predicts results from recipe and lot properties together can show which recipes are sensitive to which material, and help plan the next experiments, as in our AI formulation optimisation guide.
- Simulate the lot spread. Draw many plausible lots from the measured distribution, predict the result for each and count how often the specification is met. Liu and colleagues applied this Monte Carlo approach to variation in process settings and model error in a granulation process: they ran 50,000 simulations per point and kept the region where specifications were met with at least 90% probability. In their validation, runs inside that region met specification and runs outside did not.
- Confirm with real lots. Test the chosen recipe with the most extreme lots available, in replicates, before release.
Which control strategy fits which situation?
Robust design is one option among several, and most plants combine them.
| Strategy | How it works | When it fits | Limits |
|---|---|---|---|
| Tighter supplier specification | Narrower limits on each property | One or two critical properties the supplier can control | Higher purchase cost; more good lots rejected |
| Multivariate acceptance region | A PLS model of plant data defines which combinations of properties are acceptable (Duchesne and MacGregor) | Several correlated properties and a good history of lots and results | Needs data; harder to explain to suppliers |
| Robust recipe | Choose the recipe least sensitive to lot properties | Variation that cannot be controlled or tested quickly | May cost some performance or ingredient cost |
| Per-lot adjustment | Measure each lot on arrival and adjust dosages | A fast, reliable incoming test and a validated model | Only valid inside the range of lots already seen |
The authors of the multivariate method also point out that clear specifications can help customers widen their supplier base and accept lower-cost materials that still perform well. Per-lot adjustment matches what ICH Q8(R2) calls an adaptive process step, one that is responsive to the input materials, so that the variability of input materials can be less tightly constrained.
What should you do when a supplier changes?
Treat a new supplier or a new grade as a change to the formulation, not a purchasing detail. Test the new material in the current recipe side by side with the current material, in replicates, and compare the difference with the repeatability of the test method. If the new material sits outside the range of lots the recipe was validated with, repeat the robustness check before switching.
What we have seen at KABAS AI
Raw material variability is one of the challenges named in the ADFC (Sonatrach Group) case study: raw materials for drilling fluids can vary between batches and suppliers, which requires frequent adjustments and re-validation. In that project, Virtual Lab supports "what-if" scenarios for component substitutions and dosage adjustments, which the case study describes as particularly valuable when addressing material variability, and its Industrialize phase provides guidance for material substitutions. Material variability analysis is also part of how Virtual Lab is presented on our oil and gas solution page.
The practical lesson is simple: a model can only separate a recipe effect from a lot effect if every trial records the supplier, the lot and its measured properties. At ENAD Shymeca, as at ADFC, structuring historical formulation data came before any prediction.
If you are starting, pick one product where a supplier or lot change has caused trouble, gather the lot records and incoming tests behind it, and run the steps above on that product. The chemistry solution page shows how Virtual Lab fits formulation R&D. Related reading: design of experiments vs Bayesian optimisation, golden batch analysis for variability in production, and surrogate models for chemical processes.
Sources
- FDA, Guidance for Industry: PAT, A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance, 2004
- Varner, Influences of Cement Source and Sample of Cement Source on Compressive Strength Variability of Gravel Aggregate Concrete, Mississippi DOT, FHWA/MS-DOT-RD-13-239, 2013
- Portier et al., Continuous twin screw granulation: Impact of microcrystalline cellulose batch-to-batch variability during granulation and drying, a QbD approach, International Journal of Pharmaceutics: X, 2021
- Duchesne and MacGregor, Establishing multivariate specification regions for incoming materials, Journal of Quality Technology 36, 2004
- Paris, Duchesne and Poulin, Establishing Multivariate Specification Regions for Incoming Raw Materials Using Projection to Latent Structure Models, Frontiers in Analytical Science, 2021
- ICH Q8(R2), Pharmaceutical Development, 2009
- ICH Q7, Good Manufacturing Practice Guide for Active Pharmaceutical Ingredients, 2000
- NIST/SEMATECH e-Handbook of Statistical Methods: What are Taguchi designs?
- NIST/SEMATECH e-Handbook of Statistical Methods: Randomized block designs
- Liu et al., Optimization of critical quality attributes in continuous twin-screw wet granulation via design space validated with pilot scale experimental data, International Journal of Pharmaceutics 525, 2017
Frequently asked questions
What is the difference between a robust formulation and a design space?
A robust formulation is a recipe whose results change little when raw materials vary within their normal range. A design space, as defined in ICH Q8(R2), is the combination of material attributes and process parameters that has been shown to assure quality. Robustness is a property of the chosen recipe; the design space is the documented region around it.
Should we simply tighten the supplier specification?
Sometimes, but it has a price. Tighter univariate limits reject more good lots, which usually raises purchasing costs, and they can still accept bad lots when material properties are correlated. Tighten only the attributes shown to matter, and consider a multivariate acceptance rule or a more tolerant recipe first.
How many raw material lots do we need to test?
There is no fixed number. What matters is that the lots cover the range you expect in service, including the extremes. One published study started from 12 lots of an excipient, summarised their properties with PCA and chose the most extreme lot in each quadrant, so four lots represented the spread in the designed experiments.
Can a model adjust the recipe for each new lot?
Yes, if the lot properties that drive performance are measured on arrival and the new lot falls inside the range the model was trained on. ICH Q8(R2) describes this kind of adaptive step, which responds to the input materials. The adjusted recipe should still be confirmed by a lab test until the model has a track record.
Does a supplier certificate of analysis replace incoming testing?
Not entirely. ICH Q7, the GMP guide for active pharmaceutical ingredients, accepts certificates in place of most tests only when suppliers are evaluated, asks for full analyses on at least three batches before reducing in-house testing, and requires periodic checks of certificate reliability. Certificates also report only the supplier's specification, which may not include the attributes your recipe depends on.
Is this only a pharmaceutical problem?
No. Most published methods come from pharmaceuticals and polymers, but the same effect appears elsewhere: cement lots change concrete strength, resin lots change film quality, and drilling fluid materials vary between batches and suppliers. Any product made from raw materials that vary from lot to lot faces it.



