Roughness-Driven Fill-Tuning in Complex Systems

• The paper demonstrates that tailored interface roughness can strategically fill in gaps, modulating thermal conductivity, rheological responses, and model embedding performance.
• Methodologies include controlled interdiffusion in superlattices, engineered particle textures in suspensions, and analysis of latent space roughness for targeted model tuning.
• The results imply that optimal roughness tuning achieves performance levels beyond conventional limits, such as thermal conductivities below alloy bounds and improved predictive model accuracy.

Roughness-driven fill-tuning refers to the strategic manipulation of interface or surface roughness as a means to modulate, optimize, or homogenize a target response in complex systems. This concept has been established in disparate domains—thermal transport in superlattices, shear thickening in particle suspensions, nonlinear rheology of capillary suspensions, and the data-efficient improvement of deep material embedding spaces—linked by the physical or latent “filling-in” of gaps or frustrated regions through roughness-mediated mechanisms.

1. Foundations of Roughness-Driven Fill-Tuning

Roughness-driven fill-tuning exploits the deliberate adjustment of surface or interface roughness to “fill” regions prone to undesirable responses, such as high conductivity channels, weak frictional contacts, inadequate phonon scattering, or poorly conditioned sections of model latent spaces. The core principle is to induce new interaction pathways (physical, mechanical, or representational) at intermediate scales, distinct from either perfectly smooth (sharp) or fully mixed (random) limits.

In physical materials, this translates to partial intermixing at interfaces or the creation of multi-asperity contacts. In machine learning, this involves using latent space roughness metrics to generate data that corrects isolated pockets of poor generalization or transfer. The effect is the systematic occupation and remediation of critical regions, which “tunes” the global or aggregate response beyond what can be accomplished by simple compositional or parameter averaging (Bourrianne et al., 2020, Wilson et al., 19 Feb 2025, Yang et al., 2019, Allard et al., 2022).

2. Roughness-Tuning in Thermal Superlattices

In superlattices, roughness-driven fill-tuning is achieved by engineering the interfacial width (δ) through controlled surface interdiffusion. The goal is to tailor phonon scattering such that the cross-plane thermal conductivity (κ) drops below the “random alloy limit,” a historical lower bound for alloyed systems.

Abrupt interfaces (δ = 0) scatter low-to-mid ω phonons efficiently, whereas random alloys (large δ) scatter high ω. Introducing a finite interdiffusion width δ enables simultaneous scattering of both regimes, resulting in a non-monotonic κ(δ) with a minimum at an optimal δ_opt(L) specific to the period thickness L. Only for sufficiently thick spacers (L > L_c) does this mechanism produce κ < κ_alloy. For LJ Si/Ge analogs, δ_opt ≈ 0.3–0.6 ML (0.1–0.2 nm), and L_c ≈ 10 nm (Yang et al., 2019).

This principle generalizes to systems with high mass or moderate bond mismatch. Table 1 summarizes the impact of interface width and period on κ.

Superlattice Period (L)	Interface Width (δ)	Minimum Achievable κ/κ_alloy
Short (<10 nm)	Any	≥1
Long (≥10 nm)	δ_opt ≈ 0.1–0.2 nm	<1

The method offers a handle—orthogonal to layer thickness or alloying fraction—to achieve thermal conductivities unachievable via conventional means.

3. Roughness-Driven Shear-Thickening and Rheology

In dense suspensions, roughness-driven fill-tuning is realized by engineering the particle surface texture, which modulates the transition between continuous shear thickening (CST) and discontinuous shear thickening (DST). The macroscopic viscosity under shear rate $\dot{\gamma}$ is governed by frictional contacts whose effective coefficient $\mu$ increases with root-mean-square roughness $r_{\mathrm{rms}}$:

$$\mu(r_{\mathrm{rms}}) = \mu_0 + \alpha \frac{r_{\mathrm{rms}}}{h_{\mathrm{min}}}$$

where $\mu_0$ is the smooth-sphere baseline and $h_{\mathrm{min}}$ is the minimal gap (Bourrianne et al., 2020).

Binary mixing of hydrophilic and hydrophobic particles at fixed $r_{\mathrm{rms}}$—“fill-tuning”—generates a continuously variable mean friction:

$$\mu_{\rm mix}(x) = x\,\mu_{\rm HP} + (1-x)\,\mu_{\rm HB}$$

with $x$ the hydrophilic fraction. All onset criteria for CST and DST become tunable functions of $\mu_{\rm mix}(x)$:

$$\phi_{\rm DST}^{\rm mix}(x) \approx \frac{\phi_{\rm DST}^{\rm HP}}{x}$$

This systematic “filling” of response space enables fine control of the critical volume fractions and the rheological regime at fixed total loading—beyond what is accessible with roughness or chemistry alone.

4. Asperity Filling and Capillary Suspensions

In capillary suspensions, increasing particle roughness results in a progressive consumption of secondary liquid by surface asperities before the onset of pendular bridge formation. There exists a threshold $\phi_{\rm sec, min} \propto h/R$ (asperity height to particle radius) that must be reached before network-forming bridges develop (Allard et al., 2022). Empirical observations show that for $h\approx 84$ nm on $R =1.5$ μm particles, $\phi_{\rm sec, min}\approx 5\%$.

As asperities are filled, subsequent capillary bridges are fewer and weaker, reducing the Hertzian repulsion parameter $A$ (as determined by third-harmonic stress scaling) by up to a factor of three compared to smooth particles. Elevating $\phi_{\rm sec}$ to match plateau moduli, the yield strain in rough systems rises—attributed to enhanced multi-asperity friction. The elastic and viscous third-harmonic exponents shift from adhesion-dominated regimes toward load-control, reflecting a crossover in frictional law with roughness.

5. Roughness Metrics and Fill-Tuning in Latent Embeddings

In the context of foundation models for materials property prediction, roughness-driven fill-tuning leverages topological frustration analysis in the model’s latent embedding space (Wilson et al., 19 Feb 2025). A local surrogate similarity $s(x)$ is constructed by decoding local perturbations from a latent point $x\in\mathbb{R}^D$; the landscape $Ŝ(x)$ is interpolated and stationary points (minima, saddle points) are identified.

Large “frustration” values $F_i$ (quantifying steep similarity variation over short distances) signal rough, under-filled regions:

$$F_i = \exp\left(-\frac{\|x_{\rm min,i}-x_{\rm ts,i}\|^2}{2\ell^2}\right)\cdot[Ŝ(x_{\rm ts,i})-Ŝ(x_{\rm min,i})]$$

A “roughness field” $\mathcal{F}(x)$ is formed; its maxima yield a set of corrective points decoding to molecules with high potential to remediate embedding irregularities. Masked-token continued pretraining on only these 100 points (“fill-tuning”) yields general improvement ($\sim$1% ROC-AUC gain across tasks) at negligible cost, in stark contrast to random or task-specific fine-tuning which often degrades aggregate performance.

6. Practical Implementation and Design Guidelines

Roughness-driven fill-tuning prescribes domain-specific procedures, but unified themes emerge:

• In superlattices, select spacer thickness $L>L_c$, optimize interface roughness $\delta\approx0.3$–$0.6$ ML, and use high mass-contrast to maximize κ suppression (Yang et al., 2019).
• In suspensions, engineer $r_{\mathrm{rms}}$ and blend hydrophilic/hydrophobic surfaces to access desired CST/DST regime at a given $\phi$ (Bourrianne et al., 2020).
• In capillary networks, account for $h/R$ to determine the threshold for bridge-mediated connectivity and frictional scaling (Allard et al., 2022).
• In foundation models, analyze embedding roughness to generate fill data targeting frustration zones, followed by small data self-supervised retraining (Wilson et al., 19 Feb 2025).

The requisite data or atomistic modeling, characterization (SEM, BET, simulation), and/or algorithmic routines are modest in computational or experimental demand.

7. Limitations and Outlook

While roughness-driven fill-tuning consistently yields enhanced or more controllable responses, several system- and method-specific limitations are recognized. In physical systems, achieving precise interface roughness or robust chemical decoding (for model fill-tuning) can be challenging. The roughness metric itself may not capture domain-general deficiencies (e.g., property gradients in ML models), and translation to new modalities or architectures remains an open direction. Future pathways include iterative fill-tuning with active learning, adaptation to LoRA-style parameter-efficient updates, and cross-modal generalization beyond current model classes (Wilson et al., 19 Feb 2025).

A plausible implication is that, in both materials systems and data-driven models, roughness-driven fill-tuning operates as a kinetic or topological analog of optimal alloying: “filling in” the landscape—whether phonon, frictional, capillary, or latent space—achieves regimes of performance and tunability inaccessible to naive averaging or random mixing.