{"id":"ef9f5611-11bd-4ec3-8e71-fb5035c0ff41","arxiv_id":"1908.07628","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"PRISM-based scattering matrices simulate STEM-EELS maps with thickness-linear scaling and essentially identical accuracy to multislice, enabling an 80 Å FePt nanoparticle map in 16 hours.","lead":"A new scattering-matrix algorithm makes electron energy-loss spectroscopy (EELS) image simulation scale linearly with specimen thickness instead of quadratically, cutting compute times for large nanoscale objects from weeks to hours. The authors demonstrate an atomic-resolution EELS map of an 80 Å FePt nanoparticle in 16 hours, a calculation they estimate would take at least 80 days with conventional multislice methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inverse multislice (Eq. 19) is the step that makes PRISM-EELS linear in thickness, but its bandwidth-loss error is validated only on SrTiO3 (≤1%); the FePt demo applies it to strongly scattering Pt without a reference check, so the headline 16 h map lacks a bound on the dominant approximation.","rationale":"The reader's weakest-assumption statement identifies exactly the inverse multislice approximation as the load-bearing unvalidated step, and also flags f=9 and the transition subset. Of those, the inverse multislice is the most method-specific and is the step that converts the algorithm's scaling from quadratic to linear, so it is the right focus for a stress test. The concern is not that the method is wrong: the SrTiO3 benchmark is a real, parameter-free comparison with conventional multislice and shows sub-0.01% agreement for the baseline PRISM-EELS method, and the ~1% error from the inverse step is quantified for that system. The issue is scope: the FePt demonstration uses the same inverse operation in a regime where the stated failure mode is expected to be much larger, and no reference calculation is given. The paper's own limitation statement in Sec. IV C is correctly flagged by the reader. The suggested test is concrete and feasible: the PRISM calculation without the inverse optimization is estimated at about 2 days, which is expensive but far cheaper than the 80-day conventional multislice run, so the comparison isolates the inverse-step error without requiring an infeasible full reference. Given that the central claim is conditional on this unquantified error, the reader's CONDITIONAL verdict is appropriate and should not be changed.","tokens_in":12895,"tokens_out":9737,"duration_ms":110997,"concrete_test":"Run the Sec. IV D FePt calculation (or a strongly scattering Pt-containing subvolume of comparable thickness) twice with identical parameters: once with S2 recomputed exactly by forward multislice from each ionization slice (the Sec. IV A algorithm) and once with the inverse-multislice update of Eq. (19). Compare the resulting Fe L-edge maps using the paper's total and site-error metrics, ε_T and ε_S. If the inverse update changes site intensities by substantially more than the ~1% level seen in SrTiO3, the headline 16 h result is not a validated multislice-equivalent map and the linear-scaling claim requires a documented accuracy bound for strong scatterers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the algorithm is linear-scaling and introduces no additional error depends on the inverse multislice update of Eq. (19). Without it, the PRISM-EELS pseudocode in Sec. III recomputes S2 from scratch for each slice, costing O(nz^2); replacing that recomputation with S2 <- M^{-1} S2 is what makes the wall-clock time grow roughly linearly with thickness. The paper's own validation of this inverse step is limited to SrTiO3 in Sec. IV C, where it reports total errors around 1% (Fig. 6). That is a weak-scattering test. The FePt nanoparticle in Sec. IV D contains 16,627 Pt atoms, and the paper explicitly discusses strong high-angle scattering from Pt as producing the 'donut'/'volcano' artifacts in Fig. 7(c). The inverse operation cannot recover intensity scattered beyond the computational bandwidth, as the text acknowledges: 'a forward multislice iteration can cause electrons to scatter to high angles outside the bandwidth limit of the calculation and these electrons will not be recovered with an inverse operation.' For the FePt object the same inverse step is applied over 45 slices, with PRISM interpolation factor 9 and only 4 of the Fe L-edge transitions, and no comparison against either a full multislice reference or even a PRISM calculation without the inverse-multislice shortcut is reported. Therefore the error contribution of the single approximation that buys the linear scaling is unquantified in the headline demonstration, so the 'modest penalties' claim and the quantitative interpretation of Fig. 7 are not yet supported for strongly scattering heterogeneous specimens.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the PRISM method to STEM-EELS simulation. Instead of propagating each inelastic wave separately for every scan position and transition, the algorithm stores two scattering matrices: S1 propagates the elastic probe to the plane of ionization and S2 propagates the inelastically scattered wave to the EELS detector. This reduces the number of multislice operations and changes the runtime scaling with specimen thickness from quadratic to approximately linear. Two further optimizations are presented: evaluating Eq. (11) on a cropped real-space grid around the transition site, and an \"inverse multislice\" operation (Eq. 19) that retreats S2 by one slice rather than recomputing it from the exit surface. The base PRISM method is validated against conventional multislice for SrTiO3 with total image errors below 0.01%, the inverse-multislice approximation gives errors around 1% for SrTiO3, and the cropping optimization has error/accuracy trade-offs quantified in Fig. 5. The paper closes with an 80 Å FePt nanoparticle EELS map computed in 16 hours, compared with an estimated 80 or more days for conventional multislice, and uses the map to demonstrate a bright-field-correction strategy for interpreting STEM-EELS images.","tokens_in":13249,"tokens_out":7311,"duration_ms":280135,"significance":"If the claims hold, this is a valuable contribution: it offers a concrete path to quantitative atomic-resolution STEM-EELS simulation for heterogeneous nanostructures that are currently too expensive to treat with conventional multislice methods. The paper's strengths include an explicit pseudocode description, measured FFT-based runtime models, validation against an independent conventional multislice implementation, and supplementary MATLAB code for the core algorithms. The quantitative error metrics in Figs. 4 and 6 and the falsifiable FePt demonstration are assets that go beyond a purely algorithmic proposal. The main weakness is that the linear-scaling inverse-multislice approximation is validated only on a relatively weak-scattering SrTiO3 specimen, while the headline FePt demonstration applies the same approximation to a strongly scattering 16,627-Pt-atom object without a reference calculation, leaving the dominant approximation error in the central demonstration unquantified.","major_comments":[{"comment":"The inverse multislice approximation in Eq. (19) is the step that makes the algorithm's runtime scale linearly with thickness, but it is validated only on SrTiO3 in Fig. 6, where the errors are about 1% and the specimen is a relatively weak scatterer. The text itself notes that electrons scattered beyond the bandwidth limit are not recovered by the inverse operation. In the FePt demonstration of Sec. IV D, the same inverse operation is applied over 45 slices to an object containing 16,627 Pt atoms, which the paper identifies as producing strong high-angle scattering; no comparison is reported against either conventional multislice or PRISM without the inverse shortcut. As a result, the dominant approximation behind the headline 16-hour calculation has no quantified error bound for the object in which it is used.","section":"Sec. IV C and Sec. IV D, Eq. (19)"},{"comment":"The FePt EELS map combines an interpolation factor of f=9, a 4 Å cropping window, and only four of the Fe L-edge transitions (about 90% of the signal), yet no reference calculation is provided for this object. Since the paper claims atomic-resolution quantitative maps and interprets the artifact morphology in Fig. 7(c), the combined spatial error of these approximations should be characterized, for example by comparing a cropped subvolume or a reduced scan against a full-multislice or at least an inverse-free PRISM calculation.","section":"Sec. IV D, Fig. 7"},{"comment":"As printed, Eq. (17) does not contain any factor for the number of slices or inelastic transitions, so it does not by itself demonstrate the claimed linear scaling with nz; the text's argument that matrix multiplications dominate is plausible, but the missing sum or product over transitions should be written explicitly. The FePt runtime section reports an estimated runtime of 2 days from Eqs. (16) and (17) and an actual runtime of 16 hours without explaining the factor-of-three discrepancy; the runtime model would be more credible if this discrepancy were reconciled or the estimate revised.","section":"Sec. III, Eq. (17)"}],"minor_comments":[{"comment":"The phrase \"for of a nanoparticle\" should read \"for a nanoparticle.\"","section":"Abstract and Sec. I"},{"comment":"The text refers to \"11.4% as in Fig. 5(b)\", but the caption lists 11.5% for panel (c); the cross-reference should be corrected.","section":"Sec. IV B / Fig. 5"},{"comment":"\"Energy loses\" should be \"energy losses\" in the phrases \"for energy loses 1 eV over the ionization threshold.\"","section":"Sec. IV B and Sec. IV D"},{"comment":"The notation \"z/n z\" for the slice thickness is ambiguous; it should be written as z/n_z or Δz consistently.","section":"Eqs. (5) and (6)"},{"comment":"The typo \"In this sectio nwe\" should be corrected to \"In this section we.\"","section":"Sec. IV A"},{"comment":"The error metric should specify explicitly whether the site error εS uses the same normalization as the total error εT or a local normalization over the site window.","section":"Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of circular validation or invented entities; the validation strategy is sound as far as it goes. The revision should focus on substantiating or tempering the claim that the FePt demonstration carries the same small error as the SrTiO3 tests, and on reconciling the runtime-model discrepancy in Sec. IV D."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution is real: the PRISM scattering-matrix idea is extended to EELS in a way that avoids recomputing the post-transition propagation from scratch for every slice. Writing the inelastic transition as S2 H S1 and computing S2 via transpose multislice is a clean, correct factorization, and the base algorithm is validated carefully against conventional multislice on SrTiO3 with errors under 0.01% total and 0.02% at the atomic site. Those numbers are convincing, and the supplementary code is a plus. The runtime scaling plots track the analytic estimates well, which gives me confidence the complexity argument is honest.\n\nThe soft spot is exactly where the stress-test note lands. The inverse multislice step (Eq. 19) is what makes the wall-clock time truly linear in thickness, but its error is characterized only for SrTiO3, where total error sits around 1%. For the 80 Å FePt nanoparticle, that same inverse step is applied over 45 slices to a specimen full of strongly scattering Pt, with PRISM factor 9 and only four of the Fe L-edge transitions. No reference calculation is reported for that object. The paper does acknowledge the bandwidth-loss limitation of the inverse step in words, but it does not quantify how that loss propagates into the FePt map. So the 16-hour calculation is an impressive demonstration, but it is not yet a validated quantitative result for strong scatterers. The 'modest penalties' phrasing in the abstract applies to the SrTiO3 test, not to the FePt case.\n\nAlso worth noting, though minor: the runtime speedups depend on the FFT/multiplication balance of the specific platform, so the 'days to hours' claim should be read as representative rather than universal. The paper is honest about this, but the headline numbers in the abstract may oversimplify.\n\nOverall, the base algorithm is sound and the paper is worth a serious referee. The main requested revision should be a validation of the inverse-multislice approximation on a smaller strongly scattering object with a full multislice reference, or at least a comparison against PRISM without the inverse shortcut. That would either confirm the FePt map or reveal how much error the shortcut introduces. I would accept this for review and would cite the base algorithm if I were working on STEM-EELS simulations.","headline":"A genuinely faster STEM-EELS algorithm with a solid base validation, but the inverse-multislice shortcut that buys the linear scaling is only tested on a weak scatterer, leaving the headline FePt demo without a quantitative error bound.","tokens_in":13803,"tokens_out":1500,"would_cite":true,"duration_ms":530036,"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":"For STEM-EELS image simulation, precomputing two scattering matrices turns quadratic-in-thickness multislice cost into roughly linear scaling, matching conventional results to under 0.01% and making an 80 Å FePt nanoparticle map a 16-hour…","keywords":["scanning transmission electron microscopy","electron energy-loss spectroscopy","multislice simulation","PRISM algorithm","inelastic scattering","transition potentials","nanoparticle elemental mapping","linear-scaling algorithm"],"falsifier":"Run the same 80 Å FePt nanoparticle with a full conventional multislice STEM-EELS simulation, using all Fe transitions and no interpolation or cropping, and compare pixel-by-pixel with the 16-hour PRISM result: if the difference in high-Pt regions exceeds the ~1% site-error level seen in SrTiO3, the claim that the speedups add no meaningful error for strongly scattering objects fails.","tokens_in":12696,"feed_emoji":"🔬","tokens_out":11101,"duration_ms":102215,"temperature":0.7,"pith_summary":"Scanning transmission electron microscopy electron energy-loss spectroscopy (STEM-EELS) elemental mapping requires simulating the electron probe as it scatters many times before and after a core-level ionization event, and conventional multislice simulation cost grows quadratically with specimen thickness and scan positions. This paper shows that both expensive parts can be captured once per slice in two stored scattering matrices, one propagating the probe to the ionization depth and one propagating the inelastically scattered electrons to the EELS detector, after which every probe position and transition is just a matrix multiplication. The result is roughly linear scaling with thickness, an order-of-magnitude or larger speedup, and no added error in the test case: total image error below 0.01% for SrTiO3 from 20 to 100 Å thickness. Two further approximations, evaluating transitions on a cropped real-space grid and stepping the exit-propagation matrix backward with an inverse multislice operation, give about 1% error with additional speedup. The demonstration is an atomic-resolution Fe L-edge map of an 80 Å FePt nanoparticle computed in 16 hours, a calculation estimated to take at least 87 days with conventional multislice.","feed_headline":"Iron map of an 80 Å nanoparticle: 16 hours vs 80 days","feed_subtitle":"PRISM's scattering matrices make STEM-EELS simulation scale with thickness instead of probe count.","key_machinery":"The load-bearing object is the inelastic scattering-matrix operator $S_n = S_2 F H_{n0} F^{-1} S_1$ (Eqs. 10–11), built from PRISM's plane-wave reciprocal-space interpolated scattering matrices. $S_1$ has rows indexed by reciprocal-space points within the probe aperture and columns by real-space points; it is advanced slice by slice with the forward multislice operator. $S_2$ has rows indexed by real-space points and columns by reciprocal-space points inside the EELS aperture, and is computed by propagating plane waves backward from the detector through the transpose of the multislice operator (Eq. 14). Storing these matrices once per slice turns the per-probe cost of propagating every inelastically scattered wave to the exit surface into a matrix multiplication, which is why runtime becomes roughly linear in thickness. The inverse multislice approximation (Eq. 19) makes that scaling truly linear by reusing one $S_2$ and retreating it one slice at a time, at the cost of losing electrons scattered beyond the bandwidth limit; in the SrTiO3 tests this shows up as about 1% total error.","core_discovery":"The central claim is that a STEM-EELS image can be written as $|S_n \\psi_0|^2$ with $S_n = S_2 F H_{n0} F^{-1} S_1$, where $S_1$ is the scattering matrix from the probe-forming aperture to the plane of ionization, $H_{n0}$ is the ionization transition potential, and $S_2$ propagates the inelastically scattered wave to the EELS aperture. In the PRISM-EELS algorithm $S_1$ and $S_2$ are computed once per slice with multislice and then reused for every probe position, so the number of multislice iterations no longer grows with the size of the STEM raster. In a SrTiO3 comparison, the PRISM images agree with conventional multislice images to a normalized root-mean-square total error below 0.01% for thicknesses from 20 to 100 Å, while measured computation time grows roughly linearly with thickness rather than quadratically. For the FePt nanoparticle, a PRISM interpolation factor of 9, a cropped 4 Å transition grid, and the four dominant Fe L-edge transitions produce a 16-hour calculation; conventional multislice would take an estimated 87 days even with aggressive approximations, or 170 years without them.","pith_inferences":["The same two-scattering-matrix decomposition should transfer to other inelastic signals whose transition operators are multiplicative in real space, so the only element-specific input needed is the relevant transition potential.","The inverse multislice operator is effectively a backpropagation through the specimen; its bandwidth-loss failure mode could be useful beyond EELS, for example in exit-wave reconstruction or ptychography where the same high-angle information is missing.","The memory cost of storing the scattering matrices (about 4.4 GB for the 1836×1836 FePt example) sets a practical limit, so larger objects will likely require blockwise or streaming construction of $S_1$ and $S_2$.","Because the headline calculation keeps only four of the Fe L-edge transitions and uses interpolation factor 9, a reproducibility test that recomputes the map with all transitions at $f=1$ on a smaller object would quantify the combined spatial error of those choices."],"forward_implications":["STEM-EELS simulations of nano-scale heterogeneous objects can be run routinely on desktop or GPU hardware in hours, making quantitative comparison with experimental elemental maps practical for nanoparticles, dopants, and interfaces.","Because the multislice step is done once per slice rather than once per probe, adding scan positions or additional elemental edges to the same object becomes comparatively cheap.","The cropped transition grid and inverse multislice approximations give users a controllable accuracy-speed trade-off, with roughly 1% error for a 16–24% speedup in the tested thickness range.","The simulated FePt map reproduces the strong-scattering artifacts that complicate experimental EELS maps, and shows that dividing by a matched incoherent bright-field image raises the correlation between Fe density and EELS intensity from 0.935 to 0.976, offering a concrete correction recipe."],"supporting_citations":[{"why":"Supplies the PRISM interpolation and scattering-matrix acceleration that this paper extends to EELS.","marker":"[14]"},{"why":"Provides the conventional multislice runtime model and desktop-GPU baseline used for the speed comparisons.","marker":"[12]"},{"why":"Supplies the reconstructed atomic coordinates of the FePt nanoparticle used in the 16-hour demonstration.","marker":"[16]"},{"why":"Provides the transition-potential calculations used to produce the inelastic operators $H_{n0}$.","marker":"[10]"},{"why":"Derives the inelastic transition potential in Eq. (8) that defines $H_{n0}$.","marker":"[23]"},{"why":"Describes the numerical evaluation of the transition potential used in the implementation.","marker":"[24]"},{"why":"Introduces the bright-field division correction applied to make the simulated Fe map interpretable.","marker":"[25]"},{"why":"Supplies the multislice formalism and bandwidth-limiting implementation details behind the scattering-matrix construction.","marker":"[18]"}],"fun_headline_variants":["STEM-EELS simulation: 80 days reduced to 16 hours","Linear-scaling PRISM-EELS speeds up simulations 120x","STEM-EELS mapping: 80 Å nanoparticle in 16 hours, not 80 days","PRISM-EELS: linear scaling makes STEM-EELS simulations 100x faster","80-day STEM-EELS simulation now runs in 16 hours"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a forward multislice step can be reversed by applying the inverse of the slice transmission and propagation steps, but electrons scattered beyond the simulation grid's angular range are lost and cannot be recovered; for the strongly scattering FePt nanoparticle this inverse-multislice error was not checked against a full conventional simulation.","fun_headline_variants_meta":{"raw":{"variants":["STEM-EELS simulation: 80 days reduced to 16 hours","Linear-scaling PRISM-EELS speeds up simulations 120x","STEM-EELS mapping: 80 Å nanoparticle in 16 hours, not 80 days","PRISM-EELS: linear scaling makes STEM-EELS simulations 100x faster","80-day STEM-EELS simulation now runs in 16 hours"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000501,"raw_usage":{"total_tokens":2481,"prompt_tokens":1006,"completion_tokens":1475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":1377}},"tokens_in":622,"tokens_out":1475,"duration_ms":11857,"temperature":1.0,"reasoning_tokens":1377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:01:59.031533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 80 Å FePt nanoparticle with a full conventional multislice STEM-EELS simulation, using all Fe transitions and no interpolation or cropping, and compare pixel-by-pixel with the 16-hour PRISM result: if the difference in high-Pt regions exceeds the ~1% site-error level seen in SrTiO3, the claim that the speedups add no meaningful error for strongly scattering objects fails.","supporting_citations":[{"cited_title":"Ophus, A fast image simulation algorithm for scanning transmission electron microscopy, Advanced structural and chemical imaging 3, 13 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the PRISM interpolation and scattering-matrix acceleration that this paper extends to EELS."},{"cited_title":"Dwyer, Simulation of scanning transmission electron microscope images on desktop computers, Ultramicroscopy 110, 195 (2010)","cited_arxiv_id":null,"evidence_quote":"Provides the conventional multislice runtime model and desktop-GPU baseline used for the speed comparisons."},{"cited_title":"Yang, C.-C","cited_arxiv_id":null,"evidence_quote":"Supplies the reconstructed atomic coordinates of the FePt nanoparticle used in the 16-hour demonstration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the transition-potential calculations used to produce the inelastic operators $H_{n0}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the inelastic transition potential in Eq. (8) that defines $H_{n0}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the numerical evaluation of the transition potential used in the implementation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the bright-field division correction applied to make the simulated Fe map interpretable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multislice formalism and bandwidth-limiting implementation details behind the scattering-matrix construction."}],"review_version":1}