Bubbles form in a liquid much more easily when something is already there to grow from: a trapped nanobubble, a solid particle, or a pit in a surface. This "heterogeneous nucleation" is exploited across very different fields. Nanobubble seeds lower the overpotential of electrochemical oxygen evolution. Seeded media need less laser power to trigger cavitation. Micro-cavities on a heater let viscous fluids boil at lower temperature. Suspended catalyst particles lower the acoustic pressure needed for sonochemistry. Each paper reports the same kind of result, a threshold with seeds versus without, but each measures "onset" differently, uses different gas contents and liquids, and reports it in its own units. Nobody knows how much of the spread in reported benefits comes from the seeds themselves and how much comes from the lab.
This proposal borrows the tools that clinical evidence synthesis built for exactly this problem: pooling results from studies with inconsistent definitions. It applies them to cavitation nucleation. We will build a hand-curated, double-coded corpus of seeded-versus-unseeded threshold experiments. Each result becomes one dimensionless effect size, the log ratio of the unseeded threshold to the seeded threshold. We then fit a meta-regression in which the expected effect is set by seed geometry through the Young–Laplace relation, and study-level random effects absorb protocol differences. The central question is empirical. Does a single physically derived covariate, seed curvature, explain a meaningful share of between-study variance across four very different triggering methods? And is the remaining variance dominated by how studies define onset rather than by what they seed?
Core Claim
We claim that reported heterogeneous-nucleation benefits are only comparable across studies once two things happen. First, the outcome must be restated as a threshold ratio. Second, the operational definition of "onset" must be coded explicitly, as a moderator in its own right. The onset definition might be acoustic emission, optical scattering, visible bubbles, or a current or voltage feature. Clinical meta-analysis learned this lesson the hard way. Studies of recurrent implantation failure split on whether two or three failed cycles define the condition. Antiplatelet trials use discordant bleeding scales. PyHealth's authors (PyHealth) found that "in-hospital mortality" is operationalized so differently across EHR papers that cross-paper comparison breaks down.
Our specific prediction is that a model combining a curvature term with onset-definition moderators will do two things better than an unstructured random-effects pooled mean. It will shrink the estimated between-study variance, and it will predict held-out studies more accurately. If the curvature term adds nothing beyond generic moderators, the physics is not transferring across labs. That outcome is itself worth knowing.
What the Model Is
For experiment $i$ in study $j$, let $T^{0}_{ij}$ and $T^{s}_{ij}$ be the unseeded and seeded thresholds in whatever native unit the study uses. Depending on the triggering method, the threshold is acoustic pressure, laser intensity, heater temperature above saturation, or electrochemical supersaturation. Define the effect size $y_{ij} = \log(T^{0}_{ij}/T^{s}_{ij})$. It is unit-free and pools across methods. Young–Laplace says the pressure excess a seed must overcome scales as $2\sigma/r$, where $\sigma$ is surface tension and $r$ is seed radius. So the physically expected effect should depend on $\log(2\sigma/r)$ relative to the unseeded barrier. We fit
$$ y_{ij} = \beta_0 + \beta_1 \log\!\left(\frac{2\sigma_j}{r_{ij}}\right) + \gamma^\top z_{j} + u_{m(j)} + u_j + \varepsilon_{ij}. $$
Here $z_j$ codes protocol moderators: onset definition, dissolved-gas state, seed gas species, liquid purity, and seed surface coverage. The term $u_{m(j)}$ is a random effect for triggering method, $u_j \sim \mathcal{N}(0,\tau^2)$ is the study random effect, and $\varepsilon_{ij}$ carries the within-study sampling error reported or reconstructed from each paper.
The quantity we care about is how much $\tau^2$ and the method-level variance drop as terms are added. We also watch whether $\beta_1$ is stable in sign and magnitude when any single method is dropped.
Research Program
The work has three parts.
Build the corpus and coding manual. Two coders will independently extract threshold pairs, seed size, gas species, and onset definition. We will report inter-coder agreement. The coding manual follows the logic of PyHealth's separate "task definition" stage: we fix what counts as onset before any modeling happens, so the model is not implicitly fitted to whatever each paper chose. We deliberately keep continuous endpoints. The clinical-trial knowledge graph CTKG compressed thousands of outcome measures into 200 cluster nodes. That is a reasonable move for drug repurposing, but it would destroy exactly the quantitative threshold information this question needs.
Anchor the corpus with known cases. Several papers in the file supply anchor studies spanning the four triggering methods:
- Electrochemistry. Nanobubble seeding in water splitting, with ~200 nm seeds lowering oxygen-evolution overpotential by up to 130 mV at 1 mA/cm² (Yang et al.).
- Laser-triggered cavitation. Nanobubble-seeded photothermal cavitation with a reduced laser threshold (Zeng et al.).
- Heater-driven boiling. Micro-cavity heaters lowering the nucleation temperature (Law et al.).
- Acoustic cavitation. Particle-seeded acoustic cavitation in iron-catalyst sonochemistry (Zhang et al.).
Yang et al. also report that seeding fails when the seed gas is consumed, as with O₂ seeds during hydrogen evolution. We will code seed-gas reactivity as a moderator and treat those cases as a built-in negative control: the curvature term alone should overpredict their benefit.
Test the approach on a within-domain optimization problem. Hu et al. trained a Bayesian-optimized random forest on data from a single ultrasonic reactor. The model placed the optimum at 1200 W, while experiments showed 1200 W caused aggregation and 600 W was best. The authors attribute this to small data. We will ask whether pooling acoustic-synthesis studies with a regime-aware model would have flagged 1200 W as outside the supported, stable region.
Baselines
Do-nothing reference. Predict each held-out study's effect as the unweighted mean of the others. Any model that cannot beat this has learned nothing transferable.
Strongest statistical competitor. Standard DerSimonian–Laird random-effects meta-analysis, extended to meta-regression with generic moderators (triggering method, liquid, year). This is the tool evidence-synthesis practitioners would reach for first, and it already handles between-study variance. It will be fit on the identical extracted corpus, with the same within-study variances and the same leave-one-study-out splits. It will be given the full moderator set except the curvature term. The proposal beats it if adding $\log(2\sigma/r)$ lowers leave-one-study-out prediction error and lowers $\tau^2$, and the gain holds when each triggering method is held out in turn. No reported result was found on Lacuna for any meta-analysis of cavitation nucleation thresholds, so this baseline's value will be established by us.
Black-box ML competitor. A random forest tuned by Bayesian optimization, following the Hu et al. recipe, trained on the pooled corpus with raw covariates. It gets the same hyperparameter budget we spend on the meta-regression's model selection. For Hu et al. themselves, the file reports no quantitative prediction-error figure. The reported outcome is qualitative: the model placed the optimum at 1200 W versus an experimental optimum of 600 W. That qualitative failure is the bar the pooled models must clear on the acoustic-synthesis subset.
Physics-only baseline. Predict $y_{ij}$ from the Young–Laplace term alone, with no fitted protocol terms. This tells us how much the data add beyond textbook theory.
Key ablation. Remove the onset-definition moderator while keeping curvature. This isolates the proposal's second ingredient. We predict that dropping it inflates $\tau^2$ more than dropping any other protocol variable.
Why not just…?
Why not just run a standard random-effects meta-analysis? This is the real threat, because our model is a meta-regression. The difference lies in two choices a default meta-analysis would not make: a dimensionless threshold-ratio outcome that lets four triggering methods share one model, and a physically derived covariate in place of ad hoc moderators. If the generic-moderator baseline matches us on held-out error, then the physics adds nothing across labs. We will report that as the finding rather than defend the model.
Why not just run a controlled inter-laboratory round-robin instead of synthesizing the literature? A round-robin is the gold standard for separating lab effects, and we include a small one as prospective validation: one seed type, two or three acoustic and laser setups, and multiple onset detectors run on the same events. It cannot cover four triggering methods and dozens of seed conditions on any reasonable budget. The retrospective synthesis tells the round-robin which conditions are most informative to test.
Why not just train a flexible ML model on pooled data? Hu et al. already show that a well-tuned random forest extrapolates into a physically destructive regime when its data are thin. Pooling more studies without explicit protocol structure risks learning lab identity rather than physics. The random-forest baseline and the method-held-out splits will show directly whether that happens.
Why not just trust classical nucleation theory? Theory gives the curvature scaling but not the study-specific offsets from gas content and onset detection. The physics-only baseline measures that gap directly.
Proposed Experiments
Variance decomposition. Report $\tau^2$ and method-level variance under the nested models: null, generic moderators, plus onset definition, plus curvature. Each variance estimate will carry profile-likelihood intervals.
Held-out prediction. Two protocols. Leave-one-study-out tests transfer to a new lab. Leave-one-method-out tests whether the curvature term transfers across physics, for example predicting electrochemical studies from acoustic, laser, and heater studies.
Negative control. The model without a gas-reactivity term should overpredict the benefit for reactive seed gases, such as O₂ during hydrogen evolution. The model with that term should not.
Prospective check. Predictions for the round-robin conditions will be preregistered before measurement. Success means the observed ratios fall inside the 90% prediction intervals at close to the nominal rate, and the spread across onset detectors on identical events is comparable to the fitted onset-definition effect.
Acoustic-synthesis application. Refit on pooled ultrasonic nanoparticle studies and test whether the high-power regime is flagged as unsupported, with wide intervals or a predicted optimum below it, where Hu et al.'s single-reactor forest was not.
Clinical extension, conditional on the core results. The only clinical cavitation trial in the file, the ultrasound-microbubble chemotherapy study (Chen et al.), found large perfusion gains (Peak Enhancement +48.8%) with no significant difference in objective response rate. A single 17-patient trial is too small for this machinery, so it stays motivation, not a benchmark.
References
- PyHealth: A Deep Learning Toolkit for Healthcare Applications
- A knowledge graph of clinical trials ($$\mathop {\mathtt {CTKG}}\limits$$)
- Macroscopic bubble generation promoted by nanobubble seeds as a traceless anti-fluctuation strategy for water splitting
- Photothermal Cavitation-Driven Micromotor to Penetrate Cell Membrane
- Atomization of High-Viscosity Fluids for Aromatherapy Using Micro-heaters for Heterogeneous Bubble Nucleation
- Iron Based Catalysts Used in Water Treatment Assisted by Ultrasound: A Mini Review
- Eco-friendly production of AgNPs by ultrasound-intensified continuous method, and process evaluation via life cycle assessment and machine learning
- Clinical Study on Low-intensity Ultrasound Stimulated Microbubbles to Enhance Neoadjuvant Chemotherapy in Breast Cancer