Project summary

I request US$50,000 for nine months of independent research, with approximately 10 hours of my time per week. I will test whether monitoring uncertainty in stored thermal energy helps phase-change storage controllers recognise when their models have become unreliable.

Phase-change materials absorb or release latent heat while their temperature changes little. A temperature reading alone may therefore leave substantial uncertainty about the energy available. Existing research already provides storage models, nonlinear state observers and predictive controllers. My question is narrower: when sensors drift or storage properties change, does explicitly monitoring the spread of plausible energy states improve decisions beyond monitoring prediction residuals?

This is a computational research project. It requires no industrial partner, new laboratory installation or consortium. Its public outputs will include software, benchmark cases, protocols, results and a technical report, including negative findings.

What are this project's goals? How will you achieve them?

The goal is to establish whether a latent-energy ambiguity diagnostic supplies useful additional information for reliable control under distribution shift. I do not claim to invent phase-change storage modelling, model predictive control, state estimation, conformal prediction or runtime intervention. The proposed contribution is the diagnostic's operational definition, a controlled comparison that isolates its effect, and a reproducible benchmark.

I will build an enthalpy-based finite-volume reference model for one storage geometry and a reduced model for control. The warning diagnostic will measure the energy spread among physically admissible state/model combinations consistent with recent measurements. It is an empirical indicator, not a guarantee that every possible physical state is enclosed.

Five policies will be compared: bounded rule-based control; nominal predictive control with a nonlinear observer; fixed-margin robust control; calibrated-envelope control with residual monitoring and fallback; and the same monitored controller with an additional energy-ambiguity trigger. The last two policies will share the prediction model, calibration data and fallback settings, isolating the extra trigger.

Stress tests will cover sensor bias, latent-capacity loss, reduced heat-transfer conductance, forecast error, increased demand and sensor dropout, plus a fixed compound fault. The plan specifies 6,000 screening and 3,000 confirmatory rollouts. These are planned experiments, not completed results. Episode-level data splits and paired seeds will be used; the scoring protocol will be frozen before comparative evaluation.

Public PCM datasets are candidate sources for physical comparisons. Their archive descriptions have been reviewed; raw files, licences and compatibility still require inspection. If suitable independent measurements cannot be used, I will publish a verification-only computational benchmark and remove claims of experimental validation.

Month 1: evidence register, focused novelty review and data feasibility decision.

Month 3: verified reference model and benchmark specification.

Month 5: implemented controllers, calibration protocol and preregistration.

Month 7: scored evaluation, uncertainty analysis and failure catalogue.

Month 9: independent reproduction check, archived code, final report and manuscript or preprint.

How will you know it worked?

The primary scientific question is whether adding the ambiguity trigger improves service-temperature reliability relative to otherwise matched residual monitoring. A 25% reduction in service-temperature violations against nominal predictive control is a decision target, not a promised or achieved result.

I will also report auxiliary energy, unmet demand, operating cost, empirical uncertainty coverage, fallback use and computation time. This will expose cases where temperature performance improves only by using more backup energy or refusing demand. Calibration coverage will be tested under shifts; conformal methods will not be presented as guaranteeing coverage after assumptions fail.

The project succeeds as research if it delivers an inspectable benchmark and a defensible answer, including a null or negative answer. It does not promise deployment certification, a predetermined energy saving or an emissions reduction inferred from synthetic experiments.

How will this funding be used?

Investigator research compensation: US$36,000.

Compute allowance: US$6,000.

Independent reproducibility audit: US$3,500.

Research software and data storage: US$1,000.

Open dissemination: US$2,000.

Reproduction documentation support: US$1,500.

Total: US$50,000.

Compensation covers approximately 390 investigator-hours over nine months. These are proposed allocations, not vendor quotations. Compensation eligibility and payment arrangements must be confirmed in the recipient-specific agreement. No university overhead, travel or new equipment is assumed. Direct-cost balances do not automatically become personal compensation; the award will be used for agreed project purposes.

Who is on your team? What's your track record on similar projects?

I am Prof Abdellatif Mohammad Sadeq, PhD, a mechanical-engineering researcher affiliated with TIIAME National Research University in Uzbekistan. My background includes thermal and energy systems, numerical modelling, computational fluid dynamics and optimisation. I will lead the modelling, controller implementation, analysis and public release.

The project reserves time for studying conformal methods and funds an independent reproducibility review. No external reviewer is yet appointed. TIIAME affiliation is background information and does not imply institutional endorsement or committed university resources.

Relevant foundations include:

Vrbanc et al. (2023), PCM modelling for predictive control: https://doi.org/10.3390/en16196849

Bastida et al. (2023), nonlinear state-of-charge estimation: https://doi.org/10.1016/j.apenergy.2022.120448

Gao et al. (2025), data-corrected predictive control: https://doi.org/10.1016/j.apenergy.2025.126660

Chen et al. (2026), MPC-guided imitation learning: https://doi.org/10.1016/j.applthermaleng.2026.130741

What are the most likely causes and outcomes if this project fails?

The diagnostic may add little beyond an existing observer and residual monitor. Admissible-state sets may be too broad, too expensive to compute, or misleading when the physical model omits important behaviour. Synthetic shifts may poorly represent actual storage degradation. Public experimental data may be unsuitable, and the planned controller comparisons may exceed available computation or investigator time.

I will address these risks through an early novelty and feasibility gate, one storage geometry, bounded computational experiments and a matched diagnostic-ablation comparison. Any scope change will be documented and handled according to the grant agreement. If physical validation is infeasible, outputs will be labelled computational verification only. If the diagnostic fails, its limitations and the benchmark will still be released.

How much money have you raised in the last 12 months, and from where?

None.