--- title: ThinkPower emoji: ⚡ colorFrom: green colorTo: gray sdk: static app_file: static/index.html pinned: false --- # ThinkPower The inspection workbench pairs each selected converter schematic with an averaged DC transformer and coefficient-weighted switch, capacitor-ESR and winding resistances grouped as RFSL, plus a separate charging-loss contribution. The transformer label gives the nominal voltage conversion ratio; it does not imply physical isolation. Normalized weights, the resistance sweep and model details are collapsed to keep the comparison simple. Upload review is available in a dialog. Compare hybrid power converters through traceable equations, device-area allocation, conduction resistance, and interphase balancing. The existing NGSpice application remains at https://huggingface.co/spaces/minjiechen/XFLY with its original decks and assets. ## Run Serve the repository root using `python -m http.server 8080`, then open `http://localhost:8080/static/index.html`. The published Space serves the same entry page. Analytical calculations run in the browser; no paid hosting runtime is required. `models.py` and its Python tests provide the independent reference. The local working folder also retains the optional Docker application and simulation sources; these are not required by the static site. ## Per-unit comparison The comparison uses **Rs as its common base**. No absolute resistance, load current or output power is assigned. Inputs are rho_C=Rc/Rs, rho_L=Rl/Rs and rho_Q=1/(Cf fs Rs), with defaults 1, 1 and 20. The defaults preserve the relative loss scales of the original paper example without assigning any ohmic values. Table columns are intrinsic dimensionless coefficients: `kFET=RFET/Rs`, `kCAP=RCAP/Rc`, `kIND=RIND/Rl`, `kSSL=RSSL Cf fs`. They have different bases and are not directly additive. Weighted charts use: `rout = kIND*rho_L + hypot(kFET + kCAP*rho_C, kSSL*rho_Q)`. The state-table editor also returns kFET, kCAP and kIND, using currents normalized to Iout and areas normalized to A0. The balancing/control tab remains a separately labeled dimensional verification example. ## What is implemented - The paper's five numerical comparison modes: XFLY2P, XFLY4P, 3LBUCK, 2SBUCK, and two-phase stacked BUCK. XFLY6S and XFLY8S are catalogued with the precise reference realization and area-normalization caveats. - Interactive resistance sweep and component table, SVG/CSV/JSON exports, device-area diagrams, and XFLY6S balancing/voltage-loop frequency responses. - 2:1, 4:1, 6:1, 8:1, and 16:1 operating-point presets. These vary the supported duty model, not the graph's number of levels. Invalid domains are omitted. - XFLY-Boost and XFLY-NS topology entries, without invented loss coefficients. - Schematic uploads and a vision-assisted JSON circuit graph review. A user must explicitly verify the graph and modulation before applying a catalog model. - A generic reviewed state-table evaluator: periodic capacitor charge-balance checks, normalized RMS currents, FSL coefficients, and optimal device areas. It does not establish KCL or volt-second balance from an arbitrary drawing. ## Vision setup Enter a Hugging Face inference-enabled token in the app for each request. The browser calls `https://router.huggingface.co/v1/chat/completions` directly. There is no shared server credential in this static deployment. Tokens and images are not written to localStorage or application logs. The inference provider receives the image; its model availability and account quota apply. Default model: `Qwen/Qwen2.5-VL-72B-Instruct` (editable, availability varies). Without a token, the app explicitly reports that recognition was not performed. All deterministic comparisons and the state-table editor remain functional. ## Method Models are transcribed from the current XFLY-IEEE working manuscript, Section V-F, Table I, Figures 18–20, and its analytical verification scripts. This repository includes selected schematic assets, not the full manuscript. `Rout_est = RIND + hypot(RFET + RCAP, RSSL)`. FSL and SSL represent limits of the same charging process. They are not summed. The series charging-loss increment is `RSSL,add = hypot(RFET + RCAP, RSSL) - (RFET + RCAP)`, so `Req = RFSL + RSSL,add`. The orange inset retains the raw RSSL asymptote. Separate ideal-secondary accounting shunts represent gate and Coss losses: `Rg = Vx^2/Pg`, `Ross = Vx^2/Poss`, with `Vx = M Vin`. These do not change the per-unit Req sweep. The technology example quantifies gate loss; Coss energy needs device-transition data. Gate supply demand is tracked separately. The reference model assumes two matched flat winding currents, a common 12 A0 transistor-area budget, fixed 4:1 allocations, identical device technology, zero clamp ESR, and equal main/clamp capacitance in 2SBUCK. Switching, gate-drive, core, dead-time, interconnect and ripple losses are excluded. Conduction-efficiency estimates are not measured or total efficiencies. The XFLY2P reference coefficients use the paper's effective six-switch current model and physical split-middle area allocation. A true six-device area budget requires a separate derivation. The XFLY4P coefficients apply only to the stated four-pulse sequence and omit additional clamp recharge. ## Extend `models.py` defines the catalog and equation registry. Add a graph, gate schedule, state-current derivation, supported duty domain, device technology/rating budget, and regression cases before assigning numerical coefficients to another topology. The state-table editor supports arbitrary component counts and up to 64 states. RMS-based optimum sizing is conditional on the supplied state currents; resizing may alter current sharing, which must be recomputed in the switched model. ## Verify and publish `python -m unittest discover -p 'test_*.py' -v` `python deploy.py --check` verifies the saved Hugging Face login and current Space. `python deploy.py` uploads only the explicit application allowlist, checks the remote parent revision, and leaves the separate XFLY Space and its storage unchanged. No deployment credentials are embedded in the source or Docker image. Individual device resistance labels are generated with `python annotate_schematics.py`. The five comparison drawings use the fixed 4:1 area allocations. XFLY2P labels the combined middle switch as 2Rs; each main capacitor has ESR Rc and each winding has resistance Rl. Clamp ESR is zero in the model; Cin/Cout ESR is omitted, not asserted physically zero. Boost/NS have no assigned device sizing. Each modeled topology is paired with a proportional two-column die-area map. S/Q device numbering follows the annotated schematic. The six-switch XFLY2P map explicitly shows S2a/S2b (and Q2a/Q2b), the series physical split represented by its combined middle-switch symbol. Allocations stay fixed at the 4:1 reference. ## Supplied 180 nm technology example Source: user's `XFLY_Power_Analysis.md`, read November 16, 2026. The supplied 180 nm reference is Ron=.602 ohm, Qg=2.51 pC, Qoss=1.12 pC at W=1 mm, L=.18 um. These are user-supplied example values, not independently verified TSMC PDK specifications. No voltage-rating/PVT assertion is made. `FOMg = 1.51102e-12 ohm C`, `FOMoss = .67424e-12 ohm C`. `Rparallel = .602/Wtotal_mm`, `Rs = 12 Rparallel`. `Pg = VGS fs ng FOMg / Rparallel`, `Rg = (M Vin)^2/Pg`. At Wtotal=360 mm, VGS=3.3 V, fs=10 MHz and ng=1, Pg=29.8188 mW. FET widths follow the existing fixed 4:1 optimum; they are not resized at other ratios or rederived by the document's proposed charging-duty rule. The independent physical example uses Vin=3.3 V, Iout=2 A, flying/clamp capacitance budget=4 uF, main-cap ESR=20 mOhm and winding DCR=38 mOhm. Equal main capacitors are adopted; four-cap cells reserve equal main/clamp capacitances. Cin/Cout are outside this budget. This allocation is an assumption. Qoss is not treated as switching energy. Optional effective Eoss input is pJ per mm of total width per period, averaged across actual voltage swings and energy recovery. No default Eoss or total efficiency is fabricated. `technology.js` also derives an independent, single linear RC event model: `Revent = a^2/(2 C fs) coth[d/(2 fs Rpath C)]` for `q = a Iout/fs`. Its limits are `RSSL=a^2/(2 C fs)` and `RFSL=a^2 Rpath/d`. It is not silently substituted for a whole converter's unknown periodic RC solution. The existing paper interpolation remains labeled as an estimate. No Analog Canvas project or exports are claimed; the attached Markdown was used as analysis guidance for the existing website, not as an instruction to modify a separate schematic application or create a formal report. ## Added 2:1 family The 2:1 preset selects the added BUCK2Φ and SP-SC entries. Existing stacked BUCK is retained as a distinct eight-device benchmark; BUCK2Φ is a conventional four-device synchronous buck with two inverse-coupled windings, sized equally at 2:1. Both additions use the common 12 A0 / total-width budget, sized at 2:1 rather than the original paper cells' 4:1 reference. Widths stay fixed afterward. BUCK2Φ: M=Db, RFET=Rs/6, RCAP=0, RIND=Rl/2, RSSL=0, matched flat CCM currents. SP-SC: two interleaved four-switch cells, two equal flying capacitors, no power inductor, 50% clocks. Ron=2Rs/3 for each of eight equal devices. Each phase alternates CF between (Vin,Vout) and (Vout,ground). At ideal M=1/2: RFET=2Rs/3, RCAP=Rc/2, RIND=0, RSSL=1/(8 Cf fs), Cf per phase. Req=RSSL*coth(RSSL/RFSL), with RFSL=2Rs/3+Rc/2. Its exact linear-periodic RC increment is Req-RFSL. This is not added to another RMS charging-loss term. The physical capacitor budget is split equally between its two flying caps. Only the fixed M=.5 point is shown in the conversion-ratio sweep. For one phase, q=Iout/(4fs) per half cycle. The linear RC integral gives q²/(2Cf)*coth[1/(4fs Rpath Cf)] per event, Rpath=2Ron+Rc. Four events across two phases per period give the stated two-phase Req. Stiff ports, matched cells, constant Ron/ESR and no dead time are assumed. The source relation for one cell is also in Berkeley EECS-2021-236, Eq. (3.13): https://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/Archive/EECS-2021-236.pdf `test_two_to_one.py` independently checks event integrals and FSL/SSL limits. `build_family_schematics.py` reuses the manuscript's MOSFET and coil glyphs to build only the website's new SVGs; it does not edit or rebuild the manuscript. ## Recommendation-first interface The main view starts with 2:1 / 4:1 / 6:1 / 8:1 / 16:1 design spaces, followed by silicon and passive constraints, a ranked analytical starting point, and a shortlist with explicit exclusions. Schematics, per-unit models, die maps and derivations live in the inspection workbench. The two-phase buck uses an inverse-coupled inductor; its flat-current DCR model is unchanged. `recommend.js` screens verified operating ranges, winding/capacitor counts, positive flying-capacitance budgets and estimated nominal open-loop droop. It searches total FET width within the silicon budget, holding each model's relative allocation fixed, to minimize modeled conduction + gate power. Priority can instead minimize capacitor or winding count first. A die budget is a maximum; the tool may use less than the available transistor area. Area mode: Rparallel = RonArea / (die_mm2 * usable_fraction). The editable default RonArea is **1 mOhm mm2**, explicitly an illustrative layout assumption authorized by the user, not a verified TSMC PDK/layout value. Effective width = .602 ohm mm / Rparallel. RonQg is from the supplied 180 nm example. Both area mode and direct total-width mode are supported. Drivers, Cin/Cout and capacitor layout area are outside the FET budget; the user sets the usable fraction to reserve die for other functions. Flying/clamp caps share a separate capacitance budget under the existing equal-cap assumptions. Shortlist per-unit Req uses the common full-budget Rs=12 Rparallel, even when less transistor width is used. Inspecting a candidate loads the actual sized Rs and its physical passive ratios into the per-unit diagrams; this different reference is stated in the inspection context. Design JSON exports all inputs, scope, optimized widths, component terms, exclusions and ranked candidates. This is not a complete design optimizer: no magnetic L/ripple/saturation/volume, Coss energy, device stress/PVT, control/headroom, interconnect or startup qualification is inferred. Gate ranking assumes one charge event per physical FET per period; four-pulse counts and clamp recharge remain model limitations. No native 6:1/8:1/16:1 higher-order loss model has yet been derived. Those spaces clearly show off-reference evaluations of existing models rather than invented native topology models. Ratio grouping denotes sizing/design reference, not proof of universal optimality. `test_recommendation.py` checks the independent closed-form buck width optimum, area calibration/units, hard constraints, width caps, droop-constrained resizing, priority order and honest no-model/no-fit behavior. Topology shortlist entries are visual panels with control-free SVG previews, normalized device resistance labels (Rs, Rc, Rl), modeled losses, budgets and inspection links. Excluded modeled candidates also have panels. Detailed control sequences remain available in the inspection schematics. Candidate assets are generated by `annotate_schematics.py` from the current schematic sources; these are intentional presentation views, not revision backups. ## Coupled-inductor cancellation-point policy Recommendations and visible inspection choices now list only supported cancellation-point families, not every cell that can reach the nominal ratio. 2:1 lists coupled BUCK2Phi and stacked BUCK. 4:1 lists XFLY2P, XFLY4P, 3LBUCK and 2SBUCK under their specified balanced complementary sequences. The stacked buck still uses its existing 4:1 device-size allocation; grouping at 2:1 refers to electrical excitation rather than allocation optimality. Pure SC and off-duty buck comparisons are omitted. 6:1/8:1/16:1 have no verified cancellation-point model yet, so their visible candidate lists are empty. First-order symmetric magnetic model: Lself=Lk+Lm, M=-Lm, Ldm=Lk+2Lm. With complementary 50% winding inputs at levels 0/Vstep, phase-current ripple pp = Vstep/(4 feff Ldm). Vstep=Vin for buck cells and Vin/2 for these hybrid cells; feff=fs for two-command sequences and 2fs for the specified four-pulse / 3L sequence. Ripple percentage uses Iout/2 as base. Default Lk=50 nH, Lm=1 uH and maximum pp ripple=5% are editable illustrative inputs, not characterized TSMC magnetics. The threshold is a hard screen. The ideal summed switching ripple cancels; phase ripple is finite for finite Ldm. No measured zero-ripple result is claimed. Capacitor ripple, timing/mismatch, resistive drops and regulation-duty correction can invalidate this ideal estimate. Loss ranking retains flat-current terms and excludes ripple-loss corrections. ## Output-voltage and load inputs The requested output voltage is editable independently of Vin. Output load may be entered as current (A) or power (W). Rload = Vout/Iout = Vout^2/Pout at the requested design point. The tool displays output current and output power, load resistance in ohms, and converter output resistance Rout in mOhms for each candidate. Rout aliases the existing output-referred series loss Req; Rout/Rload is shown separately. Gate/Coss shunts are not added to this series resistance. Per-unit values use the existing common full-budget Rs reference. Choosing a ratio seeds the output-voltage field with Vin/ratio. It stays editable; an explicit mismatch is not silently corrected. A match button is available. A mismatch excludes a cancellation-point recommendation, since adjusting duty away from that point needs a new ripple/control assessment. Matching Vout to the ideal ratio is not proof of loaded regulation: the constant-current open-loop estimate is Vin/ratio - Iout Rout. Rload is the requested operating-point load value; the model does not silently replace specified current with a solved resistive-load equilibrium. The main comparison no longer features a single winning topology. All qualifying candidates appear as equally styled drawing panels beside the design inputs, with the existing priority sort, numerical results and exclusion details. The GUI operating point now has only output voltage and output power inputs. Vin is derived as selected_ratio * Vout; Iout = Pout/Vout and Rload = Vout²/Pout. Changing ratio preserves output voltage/power. The separate device-reference panel also consumes this derived operating point instead of exposing competing Vin/Iout fields. Resource and technology constraints remain editable. The numerical API retains its legacy explicit-Vin mode for previous verification cases; the GUI uses voltage_mode=output_ratio and load_mode=power. Current GUI operating-point inputs are output voltage and output current. Power is derived as Vout*Iout, Rload=Vout/Iout, and Vin=ratio*Vout. Ratio changes preserve Vout and Iout. The numerical API still supports power input for compatibility; the GUI uses load_mode=current. 3LBUCK is listed at both 2:1 and 4:1. At 2:1, a balanced flying capacitor and 50% gate duty give an ideal constant Vin/2 winding input, so the ideal winding ripple is zero even without magnetic coupling. The schematic retains coupled windings. This is distinct from complementary-excitation cancellation at 4:1. The 2:1 estimate explicitly leaves capacitor-ripple, mismatch and timing residuals unquantified; it is not a measured zero-ripple claim. The existing 4:1 device allocation is retained in the loss comparison. TI SLYT807 documents the ideal D=.5 zero-inductor-voltage point. ## Loss scoreboard The default comparison is a FIFA-inspired lineup of equal circuit columns. Switch conduction, capacitor ESR, winding DCR, incremental charging and gate power are aligned as five attributes, each with physical mW values and relative 0–99 ratings. 99 is the lowest loss in that mechanism among displayed circuits; 0 is highest, ties get 99, and a single candidate has no relative rating. These are not efficiencies or absolute technology benchmarks. Composition strips show actual fractions of the modeled total; these five terms sum to it. Charging uses Iout²(Req−RFSL), not raw RSSL. Unknown Coss/core/ripple losses are shown as unquantified rather than assigned a zero loss or score. Pros/tradeoffs highlight the largest mW difference from other candidates' mean. Circuit-card view remains available for larger schematics and detailed numbers. The design JSON includes the scorecard definitions, physical powers and scores.