Imagine a liquid that should resist slow deformation yet flow more readily when worked quickly. Specifying one viscosity would leave half the design unfinished. Shear thinning describes a fall in viscosity as shear rate increases; a flow curve records that changing response.[3]
Shengli Jiang and Michael Webb searched for 30 target curves using generative AI and fluid simulations. Initially, they varied polymer architecture and solvent affinity while keeping molecular weight and concentration fixed.[1]
Give the model an architecture to search
Their earlier TopoGNN work represents a polymer as connected beads. A line, a ring and a branched structure can contain similar numbers of units while connecting them differently. The researchers encoded those connections alongside descriptors of the whole structure, then trained a model to reconstruct the polymer, classify its architecture and predict its characteristic size in dilute solution.[2]
That produces a compact mathematical representation from which the model can generate candidate structures. The earlier study checked generated candidates with molecular simulations. It also compared structures with similar dilute-solution sizes as concentration increased: their viscosity responses separated. Controlling one property made another difference easier to investigate.[2]
In the curve-design study, a prediction model proposed candidates and simulations supplied fresh results. Five search rounds improved matches; multiple architectures could reproduce easier targets. Sampling evenly across architecture classes performed similarly to adaptive selection.[1]
Change the permitted design
Harder curves remained beyond the initial choices. Shorter chains relax faster; increasing concentration compensates for lost low-shear viscosity. Those physics-guided adjustments improved the fit. Translating the models into specific monomer chemistries and synthesising precise architectures remains separate work.[1]
A missed target can now point to a particular physical constraint that needs to change.[1]
Sources & further reading
- Generative active learning across polymer architectures and solvophobicities for targeted rheological behavior npj Computational Materials · 15 Dec 2025 · Accessed 20 Oct 2026
- Property-guided generation of complex polymer topologies using variational autoencoders npj Computational Materials · 29 Jun 2024 · Accessed 20 Oct 2026
- Rheology National Institute of Standards and Technology · Accessed 20 Oct 2026