Polymer rheology — viscosity prediction across processing temperatures

Abstract

Polymer rheology — how a melt flows under shear and temperature — is the link between molecular structure and manufacturability. We characterise Aether ^[1] performance on melt-viscosity prediction across six engineering thermoplastics over the commercially-relevant processing-temperature range, achieving mean error of 8.3% across the validated viscosity range. We describe the materials, the time-temperature superposition handling, the coupled molecular-dynamics layer for entangled-melt structure ^[2], the failure mode for highly-branched polymers, and the use of the methodology in two production manufacturing deployments.

1. Why rheology matters for manufacturing

Polymer rheology is the link between molecular structure and manufacturability. Get it right and the extrusion line runs without surging; get it wrong and you have a quarter of process debugging. The economics are real: a polymer line that surges 5% of the time wastes raw material, produces off-spec product, and degrades downstream equipment. A line that runs cleanly produces measurably more profit per unit of capital.

Conventional rheology characterisation is bench-heavy. A new polymer grade goes through rheometric measurement at multiple temperatures, the data is fit to a constitutive model (power-law, Cross, Carreau-Yasuda), and the fitted model is what the process-engineering team uses for line design. The fits are reliable but expensive in time and equipment; a new grade can take weeks to characterise fully.

We wanted to see how well a model trained on simulation could predict measured viscosity at processing conditions, using molecular architecture plus a small calibration sample rather than a full rheometric campaign. For engineering thermoplastics where the molecular structure is well-characterised, the answer is approximately yes — with documented failure modes.

2. Materials covered

Six engineering thermoplastics in the validation:

• Polypropylene (PP) — covering homopolymer, random copolymer, and impact-copolymer grades.
• Polyethylene (PE) — high-density and low-density linear grades. Long-chain branched LDPE is the documented failure mode in Section 5.
• Polyetheretherketone (PEEK) — high-performance semicrystalline thermoplastic used in aerospace and medical applications.
• Polyetherimide (PEI) — amorphous high-temperature engineering plastic.
• Polysulfone (PSU) — high-temperature amorphous resin for water-treatment membranes and medical sterilisation.
• Polyetherketoneketone (PEKK) — semicrystalline aerospace-grade thermoplastic with variable crystallinity.

Molecular-weight distributions span the commercially-relevant ranges (5 × 10⁴ to 5 × 10⁵ g/mol weight-average). Tacticity and copolymer fractions are passed as structural input. Multi-modal molecular-weight distributions (broad-distribution grades produced by Ziegler-Natta or metallocene catalysis) are supported and tested.

3. Headline numbers

Mean viscosity-prediction error across the six materials and the validated viscosity range (10⁰ to 10⁴ Pa·s, covering the typical processing window for extrusion and injection moulding) is 8.3%. This is competitive with the accuracy of bench rheometry on the same materials when the rheometric fit uses fewer than three temperatures, and within roughly 2× of a fully-characterised rheometric fit.

Per-material breakdown: PEEK and PEI are easiest (5.8% and 6.4% mean error respectively); PSU and homopolymer PP fall mid-range (7.9% and 8.1%); PE and PEKK are hardest among the six (9.5% and 9.7%). The pattern correlates with how well-represented the chain architecture is in the public materials-science literature, which itself correlates with how well-represented it was in the model's training corpus ^[1].

4. How the model handles temperature

Time-temperature superposition is the workhorse of polymer-melt characterisation. Viscosity at a given shear rate and temperature relates to viscosity at a reference temperature through a shift factor; if the shift factor is known, a single master curve at one temperature predicts behaviour at all temperatures. The shift factor in turn depends on molecular architecture and on temperature in well-characterised ways (the Williams-Landel-Ferry relationship for amorphous polymers near glass transition; Arrhenius behaviour for semicrystalline polymers above the melt range).

Aether predicts the shift factors directly from molecular architecture and a small calibration sample at a reference temperature. The customer measures the master curve at one temperature (typically the recommended processing temperature for the grade) and provides the structural input; the model extrapolates across the processing window. The temperature dependence of the shift factor is a learned function of the chain architecture rather than a fitted-per-grade parameter.

This is what makes the methodology useful operationally. A new grade is characterised at one temperature on the bench (a single afternoon of work) rather than at four or five temperatures (a week of work) before the process-engineering team has a viscosity-vs-temperature curve they can use for line design.

5. Coupled MD for melt structure

Below the entanglement molecular weight, simple viscosity-vs-MW power laws are sufficient. Above entanglement (which most engineering thermoplastics are, at commercial molecular weights), the melt has structure — chain reptation, entanglement networks, slip behaviour, sometimes long-range collective dynamics — that matters for the rheology and that simple closed-form models do not capture.

Aether calls into a molecular-dynamics agent ^[2] for melt-state structure when the molecular architecture warrants it. The MD call computes the entanglement density, the chain end-to-end distribution, and the reptation time at the reference temperature; the rheology prediction uses these structural inputs together with the architecture-dependent shift factors.

The MD call is cached and reused across temperature points. A single MD computation at the reference temperature serves the full temperature sweep; subsequent grade-variant predictions reuse the cached structure when the architecture is similar enough. This is what keeps the methodology computationally tractable — the expensive step happens once per architecture rather than once per (architecture, temperature) pair.

6. Where the model is weakest

Highly-branched polymers remain the failure mode. Long-chain branched grades (long-chain branched LDPE in particular) have rheology that depends sensitively on the exact distribution of branch points along the chain — number of branches per chain, branch length, branch-of-branch architecture. The training corpus under-represents this regime, and small differences in the branching distribution produce substantial differences in the measured viscosity that the model does not currently capture.

Mean error on long-chain branched LDPE rises to roughly 18%, which is not at production-grade accuracy. We document the failure mode explicitly to customers using branched grades; the operational response is to supplement with rheometry-based calibration on every grade, treating the model as a starting estimate rather than as a replacement for measurement.

Other failure modes are smaller but documented. Reactive blends where the polymer chemistry changes during processing are out of scope; cross-linkable grades during cure are out of scope; filled compounds where the dominant rheology is filler-matrix coupling rather than polymer chain dynamics are out of scope.

7. Manufacturing implications

Two manufacturing customers use the methodology in production for extrusion process-window optimisation. The model proposes operating conditions that hit target output without exceeding the screw's shear-stress envelope; the process-control agent maintains the conditions against drift in raw-material properties and ambient conditions ^[3].

Operational benefits, in the customers' own measurement: a 30–40% reduction in the time required to commission a new polymer grade on an existing extrusion line, and a measurable reduction (4–7%) in off-spec product rate during steady-state operation. The mechanism for the second effect is the process-control agent's ability to anticipate drift; the rheology prediction tells the controller what to expect under a small temperature or shear-rate change, and the controller pre-corrects rather than catching up to a measured deviation.

We are working on a coupled die-swell prediction module for thin-film extrusion. Die swell — the diameter expansion of polymer melt as it leaves the die — depends on melt elasticity in addition to viscosity; predicting it requires a richer rheological characterisation than what the current methodology provides. Preview results are encouraging but the module is not yet at production-grade reliability and is currently shipping in preview mode to customers who specifically request it.

8. Limitations

Three meaningful limitations. First, the methodology is calibrated on engineering thermoplastics; commodity polymers with broad branching distributions (LDPE in particular) are the documented failure mode. Second, reactive systems (cross-linkers during cure, polymer blends where chemistry changes during processing) are out of scope. Third, melt elasticity (the basis for die-swell prediction) is not currently captured at production-grade accuracy; viscosity is.

Reproducibility: the validation data for the six engineering thermoplastics is open under the Research tier where the underlying measurements are public. Some grades are proprietary to the customer and the underlying data is not shareable; aggregate per-material error statistics are.

9. Conclusion

Aether predicts melt viscosity for engineering thermoplastics within 8.3% across the processing-temperature range, using molecular architecture plus a single-temperature calibration sample. The methodology combines learned time-temperature shift factors with a coupled MD computation ^[2] for entangled-melt structure. Long-chain branched polymers remain the failure mode and require bench supplementation. Two manufacturing customers run the methodology in production for extrusion process-window optimisation, reporting 30–40% reduction in grade-commissioning time and 4–7% reduction in off-spec product during steady-state operation ^[3].