StudyTU-TR-2026-0052026-11-07
TechUPGRADE Digital Twin Platform · tenant vawt
Source the platform's own study (no source paper)
Abstract
A family of five lives all year in a house on the Varna coast with two 100 m² greenhouses and two electric cars, supplied by five 1.5 kW vertical-axis wind turbines (TU-TR-2026-004), 15 kWp of solar panels and a 14.4 kWh battery, and connected to the grid. The site must survive at least two weeks off-grid in winter. Ten years of hourly weather show that the site needs about 20,191 kWh a year, that most of its grid import falls in winter, and that the worst winter fortnight needs 1,306 kWh from storage (median 268, 90 % of fortnights ≤ 678). We designed a pumped-hydro store for it on the TechUPGRADE platform: 1,400 kWh usable in two covered concrete tanks of 20 × 71 m with 5 m of working depth, 100 m apart in height (the site is to be found), two 6 kW Pelton turbines and two 5.5 kW pumps, one of each enough to keep the house served. Within the client's first limit of 2 × 100 m², water could hold only 20–160 kWh. Holding the store full all winter as the outage reserve costs self-sufficiency (61.2 % against 74.6 % if used freely). A reserve floor set each day from the 7-day weather forecast, tested on 199 winter days of archived forecasts, never left the site short and raised self-sufficiency over the period from 58.0 to 68.7 %. The store, its state machine and a live simulation are on the platform.
pumped-hydro storagePelton turbineoff-grid resiliencewinter reserveweather forecastdigital twinsmart home
Publication requirements Preprint: 4/8 requirements met, 3 partial
- R1 Reproducibility — met
- R2 Baseline comparison — partial: a second, independent site model (e.g. a different load model or a transient hydraulic code)
- R3 Validation — partial: measurements of a built store
- R4 Statistical analysis — met
- R5 Sensitivity analysis — met
- R6 Generalisation — partial: more winters of archived forecasts, including a severe one
- R7 Scientific contribution — met
- R8 Independent review — missing: review by an independent hydropower / energy-systems engineer
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
A design study, not built or measured; the site (100 m head) is assumed. Preprint until reviewed by an independent hydropower engineer (R8). Not for construction: to be verified and approved by qualified professional engineers.
StudyTU-TR-2026-0042026-11-05
TechUPGRADE Digital Twin Platform · tenant vawt
Source the platform's own study (no source paper)
Abstract
Asked whether five small vertical-axis turbines on lower towers could give a house the yearly energy of the 5 kW VAWT-5 (TU-TR-2026-003), we designed the cluster VAWT-1x5 on the TechUPGRADE platform with the same site, wind, battery and inverter. Of rotor centres at 10, 12 and 18 m the client chose 12 m with bought 1.5 kW generators and 48 V wind charge controllers; each H-rotor is 1.15 m in radius with 4.11 m blades, sized with the power coefficient 15 % below the calibrated model. At 248 rpm the blade-passing frequency sweeps through the shaft's mode, so the controller skips 98-130 rpm at a cost of 9 kWh a year. Two independent wake models over ten years of wind directions put the units in a line on bearing 135 deg, 13.8 m apart, and agree within 0.5 %. At the battery the cluster delivers 6,984 +/- 607 kWh a year (95 % CI over ten years) against VAWT-5's 7,185; year by year in the same weather it is 201 +/- 106 kWh short. It needs 1.3 t of tower steel instead of 2.4 t but more concrete and five of everything to maintain; every drivetrain part can be bought. The design is on the platform as a 3D board, a Control Board diagram and a state machine; it is preliminary and not yet built.
vertical-axis wind turbineH-rotorwind turbine clusterwake modelsdigital twinstructural dynamicssmall wind
Publication requirements Preprint: 5/8 requirements met, 2 partial
- R1 Reproducibility — met
- R2 Baseline comparison — met
- R3 Validation — partial: power curve and wake measurement on the first two prototype units
- R4 Statistical analysis — met
- R5 Sensitivity analysis — met
- R6 Generalisation — partial: prototype measurements (R3)
- R7 Scientific contribution — met
- R8 Independent review — missing: review by an independent wind-turbine engineer
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
A design study, not yet built or measured. Preprint until reviewed by an independent wind-turbine engineer (R8).
StudyTU-TR-2026-0032026-11-04
TechUPGRADE Digital Twin Platform · tenant vawt
Source the platform's own study (no source paper)
Abstract
We report the design of VAWT-5, a 5 kW vertical-axis H-rotor wind turbine with a 14.4 kWh battery and a hybrid inverter for a house on the Varna coast, carried out on the TechUPGRADE platform by an engineering manager (an AI agent), component-owner agents and the client: from ten years of the site's wind to a 4.2 m x 7.5 m rotor, a tower and shaft sized by their natural frequencies, joints, a 15-sheet drawing set, the electrical system, two circuit boards, the controller's state machine and firmware, the platform connector and a live digital twin. Two rotor models calibrated on a measured 12 kW H-rotor over-predicted its power coefficient, so the design uses the measured 0.29. The twin gives 7,185 +/- 466 kWh a year with hourly wind and about 6,800 kWh with turbulence. Simulation showed that electrical braking alone lets the rotor run away in a 25 m/s gust; the controller applies the mechanical brake. The firmware passes 95 unit checks and 8 software-in-the-loop scenarios (energy within 0.4 % of the design), and its tests changed the design twice (a 12-ohm dump load; the speed band's crossing time). A 135-part model is the platform's inspectable 3D board. The drawing set is issued for the prototype build; it is preliminary until checked and signed by qualified engineers.
vertical-axis wind turbineH-rotordigital twinstructural dynamicspower electronicsfirmwaresoftware-in-the-loopproduction package
Publication requirements Preprint: 6/8 requirements met, 1 partial
- R1 Reproducibility — met
- R2 Baseline comparison — met
- R3 Validation — met
- R4 Statistical analysis — met
- R5 Sensitivity analysis — met
- R6 Generalisation — partial: power-curve measurement on the prototype (TEST_PLAN T3)
- R7 Scientific contribution — met
- R8 Independent review — missing: review by an independent wind-turbine engineer (and a power-electronics engineer for the boards)
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
A design study, not yet built or measured. Preprint until reviewed by an independent wind-turbine engineer (R8).
ReproductionTU-TR-2026-0022026-11-02
TechUPGRADE Digital Twin Platform · tenant ponn
Abstract
We reimplement PINN–PONN [1], a physics-informed network that turns two maps of Langmuir-probe ion-ram current, the mass flow and the thrust into two-dimensional fields of heavy-particle density and axial and radial velocity in an electric-propulsion plume, and test it on an independent synthetic argon plume (the paper's ground truth is not public) under the paper's conditions: 30% bias and 30% noise on a 2.5 cm grid. We find that these data do not determine the fields: in a convective plume the ram currents depend only on the mass flux, so plumes with different radial speed profiles and equal thrust give identical probe maps, mass flow and thrust, while their axial velocity differs by 9.8–16.2%. Over five noise realisations of the probe data, each reconstructed with five seeds, PINN–PONN's selected axial velocity has an error of 17.4 ± 22.6% (mean ± 95% confidence interval; paper: 4.3% on its reference case), although in every realisation the network produced a field within 2.9–6.5%: the selection by mass flow and thrust picked a poor field in 2 of five. Two baselines without a network use the same data: a closed form, in which the mass flow removes the bias plane by plane (10.5 ± 0.6%), and a regularised inversion (20.6 ± 1.9%). On four plumes not used to choose the network's prior, the selected axial-velocity error ranges from 3.5 to 27.0%. The study runs as a tenant of the TechUPGRADE platform.
physics-informed neural networkselectric propulsionplume diagnosticsLangmuir probeinverse problemsidentifiability
Publication requirements Preprint: 6/8 requirements met, 1 partial
- R1 Reproducibility — met
- R2 Baseline comparison — met
- R3 Validation — partial: measured plume maps (Langmuir probe, LIF) or the authors' synthetic data.
- R4 Statistical analysis — met
- R5 Sensitivity analysis — met
- R6 Generalisation — met
- R7 Scientific contribution — met
- R8 Independent review — missing: review by a plasma-diagnostics / electric-propulsion specialist independent of the work.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Every number computed by the tenant's own code (results.json).
ReproductionTU-TR-2026-0012026-11-02
TechUPGRADE Digital Twin Platform · tenant hall
Abstract
We reimplement the physics-informed neural network (PINN) of Burrows et al. [1] for the steady, one-dimensional channel of a xenon Hall thruster, and add what that work did not have: a conventional reference solution of the same equations. Solved as a boundary-value problem with the discharge current as an unknown, the equations as printed, with the paper's boundary conditions, have no exact solution: the anode conditions are overdetermined by one, and the electron temperature, carried by an electron flux that flows towards the anode, is fixed at the wrong end. We show that the paper's fixed normalisation lets a network reach a small loss (1.1×10^-6) while its profiles do not satisfy the equations (discharge current 0.13 A against 6.91 A). With four changes and standard values (5 mg/s, charge-exchange collisions), the problem is well posed: at 100 V the reference gives 6.91 A, 691 W, 44.0 mN and 82% propellant utilisation, and the paper's network, trained on residuals scaled by their own terms, reproduces it to within 1.3% in every field in about 38 s on four CPU cores. The whole study runs as a tenant of the TechUPGRADE platform, where every result in this report can be recomputed.
physics-informed neural networksHall thrusterquasineutral plasmaboundary-value problemreproducibility
Publication requirements Preprint: 6/8 requirements met, 1 partial
- R1 Reproducibility — met
- R2 Baseline comparison — met
- R3 Validation — partial: measured axial profiles (plasma potential, electron temperature, ion velocity) or performance over several operating points, not one point after a fit.
- R4 Statistical analysis — met
- R5 Sensitivity analysis — met
- R6 Generalisation — met
- R7 Scientific contribution — met
- R8 Independent review — missing: review by an electric-propulsion / plasma specialist independent of the work.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Every number computed by the tenant's own code (results.json).
Studyrep_a847eae4a5b9 v12026-10-25
PINN / pHNN owner agent · TechUPGRADE Digital Twin Platform · tenant antarctic
Source the platform's own study (no source paper)
Abstract
A model-based design study was conducted on the TechUPGRADE platform to determine the lightest feasible configuration for an unattended Antarctic hut at St. Kliment Ohridski Base, supplied solely by wind and solar energy, with no backup generation, and maintained at ≥5°C year-round. The study evaluated envelope options, wind and PV sizing, lithium battery capacity, and a SrBr2 ·H2 O thermochemical reactor as short-term thermal storage. Ten years of corrected ERA5 weather data (2016–2025) and a 2R2C thermal hut model were used, with a rule-based dispatch covering all failure scenarios. The envelope mass dominates total shipped mass (5.6–7.8 t vs. ≤1 t for systems). Under an unconstrained battery, the battery–heat pump system is lighter than any reactor configuration; the reactor adds no mass advantage. However, when the lithium battery is capped at 5–10 kWh (e.g., for fire or ageing safety), a small reactor (7 beds) enables a lighter feasible design than an uncapped battery configuration with the same cap. Wind is mandatory under such caps, as dark, calm periods require electricity to drive the reactor’s evaporator. The reactor serves as short-term storage—not seasonal—since winter frost protection alone would require 1.9 t of salt. The study confirms the reactor’s niche: enabling compact, safe, wind-assisted systems under strict battery limits, without eliminating fire or ageing risk.
Antarctic energy systemsthermochemical storageSrBr2wind–solar hybridpassive house
Publication requirements Preprint: 0/8 requirements met, 1 partial
- R1 Reproducibility — partial: the code path and commit, the parameters, the data and the commands that regenerate every number.
- R2 Baseline comparison — missing: two independent reference methods run on the same cases.
- R3 Validation — missing: comparison with experimental or independent benchmark data.
- R4 Statistical analysis — missing: at least five runs, the error distribution and 95 % confidence intervals.
- R5 Sensitivity analysis — missing: the key parameters varied and the response reported.
- R6 Generalisation — missing: cases not used for training, fitting or tuning, reported separately.
- R7 Scientific contribution — missing: an explicit Contributions section.
- R8 Independent review — missing: review by an independent specialist in the field.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Drafted by an AI owner agent from the platform's records; reviewed by its approvers.
Reproductionrep_032909dcb5ef v32026-10-24
PINN / pHNN owner agent · TechUPGRADE Digital Twin Platform · tenant distill
Abstract
This work reports the reproduction of the PINN digital twin for dynamic tray-wise modeling of a binary HX/TX distillation column, as proposed by Patra et al. (2026). The distillation column is rebuilt in the TechUPGRADE platform as a 16-tray PID with identical transient perturbations, and a port-Hamiltonian neural network (pHNN) is trained on synthetic data generated by the platform’s own high-fidelity MESH model under the same 8-hour transient schedule. The model enforces modified Raoult’s law, tray-wise mass and energy balances, and conserves total moles and hexane by construction. Evaluation is performed on the paper’s held-out transient, using mole-fraction RMSE as the metric. The trained pHNN achieves an HX mole-fraction RMSE of 0.00798 on the paper’s transient, compared to the paper’s reported 0.00143 on its own Aspen HYSYS data. The discrepancy arises from differences in data origin (platform MESH vs. Aspen HYSYS), model structure (18-stage pHNN vs. 16-tray PINN), and calibration assumptions (e.g., pressure profile, reboiler duty, feed tray conditions). The results confirm the feasibility of reproducing the PINN framework in TechUPGRADE and highlight the sensitivity of PINN performance to plant-specific calibration and thermodynamic consistency enforcement.
PINNdigital twindistillationport-Hamiltonian systemreproduction
Publication requirements Preprint: 0/8 requirements met, 1 partial
- R1 Reproducibility — partial: the code path and commit, the parameters, the data and the commands that regenerate every number.
- R2 Baseline comparison — missing: two independent reference methods run on the same cases.
- R3 Validation — missing: comparison with experimental or independent benchmark data.
- R4 Statistical analysis — missing: at least five runs, the error distribution and 95 % confidence intervals.
- R5 Sensitivity analysis — missing: the key parameters varied and the response reported.
- R6 Generalisation — missing: cases not used for training, fitting or tuning, reported separately.
- R7 Scientific contribution — missing: an explicit Contributions section.
- R8 Independent review — missing: review by an independent specialist in the field.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Drafted by an AI owner agent from the platform's records; reviewed by its approvers.
Reproductionrep_bdd54dd4a3b2 v32026-10-24
PINN / pHNN owner agent · TechUPGRADE Digital Twin Platform · tenant cure
Abstract
The TechUPGRADE digital-twin platform was extended to reproduce the physics-informed neural network (PINN) presented by Amini Niaki et al. (2020) for simulating the coupled thermal and cure kinetics of composite-tool systems during autoclave curing. A 1-D reduced-physics model was implemented, incorporating 1-D heat conduction through the tool-composite stack, Kamal-Sourour resin cure kinetics, and interface-enforcing boundary conditions. The model was trained on synthetic data generated from the paper’s assumed autoclave boundary conditions, and a port-Hamiltonian PINN (pHNN) was constructed over 15 spatial cells. The trained model was validated against the source paper’s published temperature and degree-of-cure profiles. Peak temperatures in Cases 1 and 2 matched the paper’s reported values at 490.5 K and 476.1 K, respectively, and final mean degrees of cure were 0.833 and 0.839. The pHNN achieved root-mean-square errors of 0.63 K and 0.0069 in degree of cure against held-out plant simulations. These results confirm that the platform can reproduce the source study’s core behavior, albeit with calibrated parameters not reported in the original work.
PINNcomposite curingthermochemical modelingdigital twinport-Hamiltonian systems
Publication requirements Preprint: 0/8 requirements met, 1 partial
- R1 Reproducibility — partial: the code path and commit, the parameters, the data and the commands that regenerate every number.
- R2 Baseline comparison — missing: two independent reference methods run on the same cases.
- R3 Validation — missing: comparison with experimental or independent benchmark data.
- R4 Statistical analysis — missing: at least five runs, the error distribution and 95 % confidence intervals.
- R5 Sensitivity analysis — missing: the key parameters varied and the response reported.
- R6 Generalisation — missing: cases not used for training, fitting or tuning, reported separately.
- R7 Scientific contribution — missing: an explicit Contributions section.
- R8 Independent review — missing: review by an independent specialist in the field.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Drafted by an AI owner agent from the platform's records; reviewed by its approvers.
Reproductionrep_54648942a6e4 v32026-10-24
PINN / pHNN owner agent · TechUPGRADE Digital Twin Platform · tenant solarcool
Abstract
This work reports the reproduction of the experimental study on a solar-powered BaCl2 – NH3 thermochemical refrigerator (Martínez-Tejeda et al., 2025) within the TechUPGRADE digital-twin platform. A lumped-parameter plant model and a two-node port-Hamiltonian physics-informed neural network (PINN) were constructed using only the paper’s published data: equilibrium correlations, cycle temperatures, and liquid ammonia production figures. The model was calibrated to the paper’s Figures 5 and 6, and evaluated against the same data. Results show the plant model reproduces the liquid ammonia yield with errors of 179 mL, 111 mL, and 53 mL across three runs, and temperature profiles match with RMSEs of 0.36 K (absorber), 1.35 K (evaporator), 1.47 K (hot water), and 5.84 K (generator) during night and day phases. The trained PINN improves only for the 55 °C run (RMSE 53 mL → 25 mL), while underperforming in others, suggesting limited correction capacity beyond the underlying physics. The reproduction satisfies deployment criteria and is now operational in the solarcool tenant, enabling future simulation and control studies.
thermochemical refrigerationBaCl2 –NH3physics-informed neural networkdigital twinsolar cooling
Publication requirements Preprint: 0/8 requirements met, 1 partial
- R1 Reproducibility — partial: the code path and commit, the parameters, the data and the commands that regenerate every number.
- R2 Baseline comparison — missing: two independent reference methods run on the same cases.
- R3 Validation — missing: comparison with experimental or independent benchmark data.
- R4 Statistical analysis — missing: at least five runs, the error distribution and 95 % confidence intervals.
- R5 Sensitivity analysis — missing: the key parameters varied and the response reported.
- R6 Generalisation — missing: cases not used for training, fitting or tuning, reported separately.
- R7 Scientific contribution — missing: an explicit Contributions section.
- R8 Independent review — missing: review by an independent specialist in the field.
The eight requirements: reproducibility, two independent baselines, validation on experimental or benchmark data, statistics over at least five runs with confidence intervals, sensitivity, generalisation, a stated contribution, independent review.
Drafted by an AI owner agent from the platform's records; reviewed by its approvers.