{"id":"1722e459-5f73-4e16-865a-32cc62f27353","arxiv_id":"2505.11236","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":26,"one_line_summary":"ForgetMeNot is a fab-calibrated model that estimates fluorinated 'forever chemical' emissions from semiconductor manufacturing, and the analysis shows these can exceed embodied carbon for modern CPUs.","lead":"This paper introduces ForgetMeNot, a modeling tool that estimates fluorinated 'forever chemical' emissions from semiconductor manufacturing using fab-specific practices and hardware specifications. It shows that these emissions can exceed embodied carbon for CPUs and can differ across vendors, nodes, and storage choices, which matters for sustainable hardware decisions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'within 5%' validation is in-sample: the TRI records used to calibrate the usage coefficients in Sec. 4.3 are reused as the ground truth in Sec. 4.4, so the headline accuracy claim is not yet supported.","rationale":"The reader's weakest_assumption already identifies the core problem: the same TRI data are used both to calibrate the usage coefficients and as the validation target. My analysis confirms this is the most load-bearing weakness, because the paper's headline contribution is the quantitative 'within 5%' accuracy claim and the claimed ability to generalize across facilities. The defect is not a disagreement with external consensus; it is an internal validation-logic problem: Sec. 4.3 derives k coefficients from TRI totals, and Sec. 4.4 reports agreement with those same TRI totals. With many free parameters, fitting four facility aggregates to within 5% is not evidence that the model would predict a new facility or a future hardware generation. A leave-one-facility-out test would settle whether the model generalizes or merely memorizes each facility's total. The qualitative findings—that fluorinated emissions can dominate embodied carbon and that hardware choices shift when they are included—remain credible and useful, but they do not depend on the 5% validation claim. Since the reader's CONDITIONAL verdict already requires out-of-sample validation and uncertainty analysis, no verdict change is needed; the condition should be stated as a hard requirement for accepting the central accuracy claim.","tokens_in":25013,"tokens_out":5636,"duration_ms":61536,"concrete_test":"Leave-one-facility-out cross-validation: recalibrate all source usage coefficients and scaling parameters using TRI-derived totals from only three of the four facilities (Intel Oregon, AMD Dresden, Samsung Texas, Micron Idaho), then predict the held-out facility's annual fluorinated-compound release for the corresponding year and product mix. If any held-out prediction error exceeds 5%, or the mean absolute error across the four held-out folds exceeds 5%, the 'within 5%' claim is an artifact of in-sample calibration rather than evidence of generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.3 computes facility-specific base usage coefficients (kEtch, kClean, kPhoto, etc.) from EPA TRI emissions after subtracting a 10-mile-radius neighborhood baseline. Sec. 4.4 then validates the model by comparing it with \"the emissions value from the TRI dataset during the time when the hardware being evaluated was being manufactured.\" The validation target is therefore the same data used for calibration. With 26 parameters and only four facility-level totals, agreement within 4.7% is a goodness-of-fit statement, not an out-of-sample accuracy result. The load-bearing premise that a TRI facility total, after baseline subtraction, equals emissions attributable to the specific modeled hardware is also unverified: fabs produce many product families and nodes simultaneously, while the model predicts per-component emissions, and the 10-mile subtraction cannot separate fab process emissions from other local sources or from other products. Many parameters are stated as \"typically\" values without uncertainty (e.g., all alpha exponents and phi_Lith in Table 1), so the 5% claim carries no error bars. Thus the central claim that ForgetMeNot \"captures emissions within 5% of measured values across facilities\" is not established by the experiments as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ForgetMeNot introduces a bottom-up model of fluorinated-compound (PFAS/HFC/etc.) emissions from semiconductor manufacturing for computing hardware. The model decomposes emissions into twelve sources (etching, chamber cleaning, photolithography, heat transfer fluids, solvents, dielectrics, wafer thinning, testing, soldering, vacuum pumps, pulsed-plasma nanocoatings, packaging), with usage per source expressed as base coefficients times die-area/step-count scalings that depend on technology node and lithography type (Table 1). Facility-specific base coefficients are calibrated from EPA TRI facility totals after subtracting a 10-mile-radius neighborhood baseline (Sec. 4.3), and the model is then validated against TRI-derived facility emissions for an Intel CPU, an AMD CPU, Samsung DRAM, and a Micron SSD, with a claimed maximum difference of 4.7% (Sec. 4.4, Fig. 9). The paper then uses the model to compare fluorinated-compound footprints across CPU generations, vendors, DRAM generations, and storage types (Sec. 5), demonstrates fab-level reduction levers (Sec. 4.2), and assembles low-emission server configurations that trade performance against manufacturing emissions (Sec. 6). The artifact is open-sourced at Zenodo.","tokens_in":25391,"tokens_out":2751,"duration_ms":30385,"significance":"If the validation claim were established, ForgetMeNot would be a genuinely useful first systems-level accounting of fluorinated emissions for computing hardware, complementing embodied-carbon tools such as ACT. The paper is also commendable for open-sourcing the tool, for decomposing emissions by manufacturing source, and for showing that including fluorinated compounds can change which hardware is ranked as most sustainable. However, the central accuracy claim currently rests on an in-sample comparison: the same TRI facility totals used to fit the usage coefficients are reused as the validation ground truth. With 26 free parameters and only four facility-level totals, agreement within 4.7% is a goodness-of-fit statement rather than an out-of-sample accuracy result. The load-bearing attribution premise, that a TRI facility total minus a neighborhood baseline equals emissions from the specific modeled hardware, is also unverified. These issues make the headline 'within 5%' claim unsupported as presented, although the modeling framework itself is plausible and worth developing.","major_comments":[{"comment":"The validation is in-sample. Sec. 4.3 derives facility-specific base usage coefficients (kEtch, kClean, kPhoto, etc.) from TRI facility emissions after baseline subtraction, and Sec. 4.4 then compares ForgetMeNot's output with 'the emissions value from the TRI dataset during the time when the hardware being evaluated was being manufactured.' With approximately 26 parameters and only four facility-level totals, a maximum difference of 4.7% does not establish that the model 'captures emissions within 5% of measured values across facilities.' I ask for an out-of-sample evaluation: for example, calibrate on a subset of facilities or years and validate on held-out facilities or years, or compare against independent, facility-reported fluorinated emissions (e.g., GHGRP electronics-manufacturing data) rather than the same TRI records used for calibration.","section":"Sec. 4.3 vs Sec. 4.4, Fig. 9"},{"comment":"The attribution from a facility total to a specific hardware component is not established. A fab typically manufactures many product families and technology nodes simultaneously, and the 10-mile-radius subtraction cannot separate fab process emissions from other local sources or from emissions attributable to other products manufactured in the same facility and time window. Moreover, Fig. 9 (left) labels per-source values as 'Actual Emissions from Fabrication Facility,' but TRI reports facility-level totals, not per-source breakdowns, so the source-level comparison cannot be supported by the cited data. Please provide evidence that the residual TRI value is attributable to the modeled component, or present the validation at the facility level with a clear description of what the TRI residual can and cannot identify.","section":"Sec. 4.3, 10-mile baseline; Fig. 9 (left)"},{"comment":"Several load-bearing parameters are hand-set or stated as 'typically' values without uncertainties: all alpha exponents, phi_Lith = 0.8, the 1.25x lower-GWP cleaning-step multiplier, and the base coefficients' typical magnitudes. Because these parameters directly determine the emissions computed for new hardware (and hence all conclusions in Secs. 5 and 6), the paper needs a sensitivity analysis: for example, vary alpha_etch, alpha_clean, alpha_photo, alpha_time, and phi_Lith over plausible ranges and report how the 'within 5%' claim and the server-ranking conclusions change. Without such an analysis, the 5% claim carries no error bars and the qualitative conclusions in Sec. 6 cannot be assessed for robustness.","section":"Table 1 and Sec. 4.1"},{"comment":"The claim that ForgetMeNot 'can effectively generalize modeling across different fabrication facilities and hardware types' is supported by only four data points (Intel Oregon, AMD Dresden, Samsung Texas, Micron Idaho), each corresponding to a different hardware type and facility. This is too few points to establish generalization, especially because the validation target is also the calibration target. Please report the number of facilities and hardware types, provide per-facility parameter fits and their uncertainties, and, if possible, include more facilities or a temporal hold-out so that the generalization claim is statistically meaningful.","section":"Sec. 4.4 and Fig. 9 (right)"}],"minor_comments":[{"comment":"The sentence in Sec. 2 stating that fluorinated compound emissions have 'consistently been 6 to 7 orders of magnitude higher than embodied carbon emissions' appears inconsistent with Fig. 3, where both quantities are plotted on the same scale in MMTCE; please check the units and the intended comparison.","section":"Fig. 3 and Sec. 2"},{"comment":"The word 'Foorptint' appears in the captions of Fig. 3 and Fig. 10; it should be 'Footprint.'","section":"Captions, Figs. 3 and 10"},{"comment":"The statement that reducing cores from 32 to 22 while increasing cache from 48 MB to 64 MB 'maintains similar performance (0.1 TFLOPs)' is asserted without a benchmark reference or a model; please provide the basis for the performance equivalence or soften the claim.","section":"Sec. 4.2, Fig. 7"},{"comment":"The paper would be easier to reproduce if Table 1 (or an appendix) listed the exact numeric values used for each k coefficient, each alpha exponent, phi_Lith, eta_rec, yield, gamma_usable, and the reference-hardware quantities for each of the four validated facilities, rather than only 'typically' values in the text.","section":"Sec. 4.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important and timely gap for the systems-sustainability community, and the open-source artifact is a strength. The main risk is methodological: the headline 5% validation is in-sample and the facility-to-hardware attribution is unverified. I believe this is fixable within the manuscript's scope by reframing the claims and adding out-of-sample or uncertainty analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: ForgetMeNot is the first systems-level model I know of that tries to quantify fluorinated compound (forever chemical) emissions from computing hardware manufacturing, and that alone makes it worth a serious referee. The authors also ship code and a calibration recipe using EPA TRI data. But the headline accuracy claim — 'within 5% of measured values' — is not supported by the validation as presented, because the same TRI facility records are used to fit the base usage coefficients and then serve as the ground truth. That is in-sample agreement, not independent validation.\n\nWhat's genuinely new: the integrated view of etching, chamber cleaning, lithography, HTFs, and other sources mapped onto CPUs, DRAM, and storage, with a public tool. The qualitative result that fluorinated emissions can exceed embodied carbon by orders of magnitude in electronics manufacturing is important and consistent with EPA data. The server assembly analysis and the GWP robustness check are useful, even if the numbers inherit the model's parameters.\n\nSoft spots: (1) The 5% validation is circular, as the stress-test note says. Section 4.3 calibrates kEtch, kClean, etc. from TRI facility totals minus a 10-mile neighborhood baseline; Section 4.4 validates against those same TRI totals. With 26 free parameters and four facility-level data points, 4.7% agreement is curve fitting. (2) The premise that a facility's TRI total, after baseline subtraction, equals emissions from the specific hardware modeled is unverified. Fabs run many product families; per-component attribution is an assumption. (3) Many parameters (alpha exponents, phi_Lith, and the lower-GWP cleaning multiplier) are hand-set \"typically\" values with no sensitivity analysis or error bars. For a model meant to predict emissions for novel hardware, the uncertainty should be quantified.\n\nI don't think the central idea collapses; the qualitative contribution — fluorinated compounds matter and hardware choices change when you include them — stands. But the paper sells the validation too hard. A revision should present the 5% claim as calibration fit, add out-of-sample or at least leave-one-facility-out validation, and show sensitivity over the hand-set exponents and coefficients.\n\nWho it's for: systems sustainability researchers, datacenter operators, and anyone using ACT-style embodied carbon models who wants the fluorine axis. It deserves peer review; the tool and data are reproducible, and the question is timely. Recommend: accept with major revision, with the accuracy claim reframed and validation strengthened.","headline":"First systems-level tool for fluorinated emissions in computing hardware, with open code and a timely qualitative message, but the 'within 5%' claim is in-sample and needs reframing.","tokens_in":25918,"tokens_out":1677,"would_cite":true,"duration_ms":17059,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ForgetMeNot: a systems-level model of fluorinated 'forever chemical' emissions in chip manufacturing, claimed accurate to within 5% of real fab emission records.","keywords":["sustainable computing","forever chemicals","PFAS","fluorinated compounds","semiconductor manufacturing","emission modeling","embodied carbon","datacenter servers"],"falsifier":"A decisive check would be direct process-level metering in one fab for one product line: measuring the mass of each fluorinated gas (NF3, SF6, CF4, C2F6, CHF3, C3F8, C4F8) consumed and captured across etching, chamber cleaning, and lithography, and comparing those source-level quantities to ForgetMeNot's per-source estimates. If the source-level estimates miss the metered values systematically while the facility total still lands within 5%, the model's internal scaling would be falsified even though its headline accuracy survived.","tokens_in":24789,"feed_emoji":"🧪","tokens_out":8565,"duration_ms":79031,"temperature":0.7,"pith_summary":"The paper argues that the sustainability conversation about computing hardware is incomplete when it counts carbon dioxide only, because the manufacturing process releases fluorinated compounds—'forever chemicals' with global warming potentials thousands of times that of carbon dioxide. To close the gap it introduces ForgetMeNot, a modeling tool that estimates fluorinated emissions from twelve manufacturing sources by combining fabrication-facility practices (gas recovery factors, wafer yield, usage coefficients) with the specifications of the hardware being made (die area, core and cache counts, TDP, node size, lithography type). The paper's central claim is that ForgetMeNot matches real facility-reported emissions within 5% across four hardware types and three different facilities, and therefore generalizes across fabs and components. If the claim holds, it gives hardware designers and datacenter operators a practical way to see and reduce the fluorinated part of manufacturing emissions rather than treating them as invisible.","feed_headline":"Chip emissions tool tracks forever chemicals within 5% of fab data","feed_subtitle":"Fluorinated compounds in hardware manufacturing can outweigh embodied carbon; a new model makes them visible.","key_machinery":"The load-bearing identity is the per-source emission formula $$E_i = N_{\\text{Wafers}} \\times \\text{Usage}_i \\times (1-\\eta_{\\text{Rec}}) \\times \\text{GWP}_i,$$ summed over twelve sources. $N_{\\text{Wafers}}$ is derived geometrically from die area, usable wafer area, wafer diameter, and yield, so that larger dies or lower yields mean more wafers and more emissions. Each source's usage is not a free constant: it is anchored to a reference older-generation component and scaled by node-size exponents ($\\alpha$), die area, TDP, package size, and a lithography factor $\\phi_{\\text{Lith}}$, which is the mechanism that makes EUV lithography less emissive than DUV. The gas recovery factor $\\eta_{\\text{Rec}}$ (about 0.9 at advanced facilities) appears in every source, which is why the paper treats small improvements in capture as a first-order lever for emission reduction.","core_discovery":"ForgetMeNot's central claim is that the fluorinated compound emissions of manufacturing a piece of computing hardware can be predicted from a small set of hardware specifications and fab-level parameters, and that the prediction matches what fabrication facilities actually report. The model decomposes the manufacturing pipeline into twelve emission sources—etching, chamber cleaning, photolithography, heat-transfer fluids, solvent fluids, dielectric fluids, wafer thinning, testing, soldering, vacuum pumps, pulsed-plasma nanocoatings, and packaging—and computes each source's contribution as wafers used times per-wafer fluorinated compound usage times the fraction not captured by gas recovery times the global warming potential of the gases involved. Per-wafer usage is scaled from a reference older-generation component using the ratio of node sizes, die-area terms for cores and cache, TDP, package size, and a lithography factor that makes EUV patterning cheaper than DUV. The paper validates the summed model against public facility-level emission records during the production of four real products—a server CPU fabricated in Oregon, a competing CPU fabricated in Dresden, a DRAM module fabricated in Texas, and a solid-state drive fabricated in Idaho—and reports a maximum difference of 4.7% between modeled and reported total emissions. On that basis it claims ForgetMeNot generalizes across different fabrication facilities and hardware types.","pith_inferences":["Because the model is calibrated on facility-wide reported totals, its twelve-way breakdown by source is an allocation based on process recipes rather than a direct measurement; the aggregate 5% accuracy does not by itself verify the per-source split.","The same parameterized structure could be extended to forecast emissions for process changes not yet in production, such as high-NA EUV or new lower-GWP cleaning gases, as long as step-count and reactivity data are available.","The dependence of optimal hardware choices on the GWP time horizon is a sensitivity worth probing: the paper notes the lowest-emission storage-optimized configuration under a 500-year horizon becomes tenth-lowest under a 20-year horizon, so procurement rankings may shift with climate-policy timeframes.","If fabs ever publish source-level gas metering, the model's exponents and per-source coefficients could be fitted independently instead of through aggregate totals, turning validation from a single number into twelve testable predictions."],"forward_implications":["Smaller manufacturing nodes increase fluorinated emissions: the paper reports a 17.5% rise moving from 10 nm to 7 nm and a further 26.9% rise moving from 7 nm to 5 nm for a high-end server CPU made with EUV lithography.","Switching from DUV to EUV lithography lowers fluorinated emissions enough that a later CPU generation can have a smaller fluorinated footprint than its predecessor, even as embodied carbon keeps rising.","DDR5 memory modules have lower fluorinated compound and embodied carbon footprints per gigabyte than DDR4 modules of the same capacity, across multiple vendors.","For storage, per-terabyte fluorinated and carbon footprints fall by roughly 25–40% as capacity increases, and SATA HDDs have substantially lower manufacturing footprints than SSDs.","Datacenter operators choosing lowest-emission components can cut manufacturing emissions by roughly 51–55% for general-purpose, compute-optimized, and memory-optimized servers compared with median configurations, while the fastest configurations buy about 8.5% more throughput at about 52% more manufacturing emissions in the general-purpose case."],"supporting_citations":[{"why":"Supplies the public facility-level emission records used to calibrate usage coefficients and to validate the model's totals against real fabs.","marker":"[15]"},{"why":"Identifies fluorinated compounds as a significant missing component of computing's manufacturing footprint, the gap ForgetMeNot fills.","marker":"[36]"},{"why":"The embodied-carbon modeling tool that ForgetMeNot complements; its carbon-only estimates serve as the comparison baseline for total emission rankings.","marker":"[43]"},{"why":"Provides the global warming potential values that convert each fluorinated gas's mass into CO2-equivalent emissions.","marker":"[41]"},{"why":"Provides the electronics-manufacturing emission series comparing fluorinated compounds with embodied carbon over thirteen years.","marker":"[4]"},{"why":"Fab process simulation used to set reference step counts and process parameters for the reference older-generation hardware.","marker":"[77, 78]"},{"why":"Supplies supply-chain process and tool data used to set realistic inputs for equipment use and emissions.","marker":"[3]"},{"why":"Source studies behind the per-source emission equation form: wafers, gas usage, recovery factor, and GWP.","marker":"[48, 67, 69]"}],"fun_headline_variants":["Chip fabs get a model for tracking forever chemical emissions","New tool quantifies fluorinated compounds in chip manufacturing","Beyond carbon: modeling fluorinated emissions from chip fabs","Forever chemical emissions from chipmaking now measurable within 5%","ForgetMeNot: a model for chip fabs' fluorinated gas footprint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy claim rests on the assumption that public facility-level emission reports, after subtracting a 10-mile-radius neighborhood baseline, truly measure emissions from the specific hardware products being modeled—even though the same reports were used to fit the model's usage coefficients.","fun_headline_variants_meta":{"raw":{"variants":["Chip fabs get a model for tracking forever chemical emissions","New tool quantifies fluorinated compounds in chip manufacturing","Beyond carbon: modeling fluorinated emissions from chip fabs","Forever chemical emissions from chipmaking now measurable within 5%","ForgetMeNot: a model for chip fabs' fluorinated gas footprint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1756,"prompt_tokens":1021,"completion_tokens":735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":647}},"tokens_in":637,"tokens_out":735,"duration_ms":7631,"temperature":1.0,"reasoning_tokens":647,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:55:09.375975+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be direct process-level metering in one fab for one product line: measuring the mass of each fluorinated gas (NF3, SF6, CF4, C2F6, CHF3, C3F8, C4F8) consumed and captured across etching, chamber cleaning, and lithography, and comparing those source-level quantities to ForgetMeNot's per-source estimates. If the source-level estimates miss the metered values systematically while the facility total still lands within 5%, the model's internal scaling would be falsified even though its headline accuracy survived.","supporting_citations":[{"cited_title":"https://www.epa.gov/toxics-release-inventory-tri-program/tri-toolbox","cited_arxiv_id":null,"evidence_quote":"Supplies the public facility-level emission records used to calibrate usage coefficients and to validate the model's totals against real fabs."},{"cited_title":"The environmental impact of forever chemicals in computing systems","cited_arxiv_id":null,"evidence_quote":"Identifies fluorinated compounds as a significant missing component of computing's manufacturing footprint, the gap ForgetMeNot fills."},{"cited_title":"Act: Designing sustainable computer systems with an architectural carbon modeling tool","cited_arxiv_id":null,"evidence_quote":"The embodied-carbon modeling tool that ForgetMeNot complements; its carbon-only estimates serve as the comparison baseline for total emission rankings."},{"cited_title":"Global Warming Potential Values (August 2024), 2024","cited_arxiv_id":null,"evidence_quote":"Provides the global warming potential values that convert each fluorinated gas's mass into CO2-equivalent emissions."},{"cited_title":"https://www.epa.gov/ghgreporting/ghgrp-electronics- manufacturing","cited_arxiv_id":null,"evidence_quote":"Provides the electronics-manufacturing emission series comparing fluorinated compounds with embodied carbon over thirteen years."}],"review_version":1}