Neurocomputing Submission Guide: How to Submit to Neurocomputing (Elsevier)

Neurocomputing publishes Research Papers, Review Articles, and Software Publications. Most submissions are Research Papers, and the journal does not enforce a single hard word limit the way a letters journal does. Length is governed by completeness: an over-long paper is judged on whether every figure, table, and ablation earns its space, not on a fixed cap. That freedom is a trap for authors who pad a thin contribution to look substantial.

Research Papers should present a fundamental contribution to neural networks or learning systems. The methodological claim has to be visible in the abstract and the introduction, because that is what an Associate Editor reads first. A typical strong submission runs in the range of 6,000 to 9,000 words with the figures and tables a reviewer needs to judge the claim, but the number that matters is the ratio of novelty to length, not the length itself.

Review Articles are expected to synthesize a subfield with a genuine organizing argument, not to list papers. A survey that catalogs methods without a thesis about where the field is going is the most common reason a review is returned.

Software Publications disseminate non-trivial implementations: toolboxes, libraries, or systems with a real algorithmic contribution. A wrapper around an existing framework is not a Software Publication.

For files, Elsevier accepts standard manuscript formats and provides a LaTeX template; Neurocomputing uses single-anonymized peer review, so author names appear on the manuscript. There is no fixed figure cap, but every figure has to earn its place: a results section carrying 12 figures and 20 supplementary panels usually signals that the main contribution is not clear enough to stand on fewer.

Keep individual figure files under the portal's per-file size limit (commonly around 10 MB) and supply vector or high-resolution sources so production does not bounce them. Manuscripts that are unclear because of English-language quality can be returned for rewrite before review, so the language bar is enforced at triage, not deferred.

Before the contribution claim and declarations are locked, a Neurocomputing learning-systems contribution check can confirm whether the methodological novelty is stated clearly enough to survive the Associate-Editor screen.

Neurocomputing assigns submissions to an Associate Editor who handles them through Editorial Manager. Community-reported data puts the first review round at roughly 1.7 months, with about 2.0 reports per submission and around 1.8 review rounds before a final decision. Treat the stages below as planning ranges, not commitments.

• Day 0: Submission and PDF build. Editorial Manager ingests your files and builds a merged PDF. You proof it, confirm the data availability statement, the competing-interests declaration, the CRediT contributions, and the highlights, then submit.

• Days 1 to 12: Editorial screening. Editorial staff and a handling editor check scope fit, format compliance, language quality, and whether the contribution is a learning-systems advance or an application.

The fastest returns happen in this window: generic deep-learning applications, out-of-scope work, and language-quality returns rarely reach external review. Community data reports immediate rejections in about 12 days.

• Days 12 to 21: Associate Editor assignment. An Associate Editor in the relevant neural-networks area takes the manuscript and decides whether to send it for external review.

Benchmark-only papers with no ablations and no analysis of why the method works are commonly returned here, before reviewers are invited.

• Days 21 to 50: Peer review. Reviewers are invited and reports return, typically two reports, on a multi-week cadence. Community data suggests a first round near 1.7 months is common, though subfield and reviewer load shift this.

• Weeks 7 to 12: Decision and revision. Reject, major revision, minor revision, or accept.

A revised manuscript must be accompanied by a point-by-point response letter. With about 1.8 rounds on average, most papers that pass review go through at least one major-revision cycle.

• Months 3 to 4: Final decision and production. Total handling time for accepted manuscripts runs to roughly 3.1 months from submission, with faster outcomes for clean, well-scoped contributions and slower ones for papers that need a second review round to settle reproducibility questions.

In our pre-submission review work with Neurocomputing manuscripts, four patterns generate the most consistent early returns. None of them are about the model being wrong. They are about whether the work makes a neural-networks or learning-systems contribution the field can build on, which is exactly what this journal screens for before peer review begins.

In our review of neural-networks manuscripts, each of these is a named rejection pattern you can check your own draft against, and each reflects an editorial triage pattern specific to how Associate Editors at this journal read learning-systems submissions. Because the editor decides from the abstract, the figures, and the methods rather than from a pitch, a weak contribution claim has nothing to hide behind.

Editors consistently screen for these before sending a manuscript out for review.

Neurocomputing's Guide for Authors and Editorial Manager requirements define the mechanics below; the patterns describe how manuscripts coming through pre-submission review for this journal most often fall short of them. SciRev community data on this journal, where authors report a first review round near 1.7 months and immediate rejections in about 12 days, is consistent with what we see: most attrition happens at the Associate-Editor screen, before reviewers ever weigh in, and these four patterns are why.

A generic deep-learning application with no neural-network or learning-systems contribution. The single most common stall we see is a paper whose real story is "we applied a standard architecture to a new dataset." The model is a known transformer, CNN, or LSTM, the dataset is the only new element, and the methods section describes configuration rather than a learning-systems advance.

An Associate Editor at Neurocomputing reads the abstract and asks the obvious question: what does this teach the field about neural networks or learning that it did not already know? When the answer is "this domain now has a baseline," the work is an application paper, and Neurocomputing returns it because the contribution belongs in a domain venue, not a neural-computing journal.

The fix is to make the architecture, learning rule, or training analysis the protagonist, not the dataset.

Check whether your Neurocomputing paper makes a real learning-systems contribution →

Benchmark-only results that report a small gain with no ablations or analysis. The parallel failure on the evidence side is a results section built entirely around a leaderboard. The paper reports a one- or two-point accuracy improvement over prior methods on standard benchmarks, but there is no ablation isolating which component drives the gain, no analysis of why the method works, and no sensitivity study showing the result is not a tuning artifact.

Reviewers in this field treat ablations as part of the result, not an appendix, so a table of accuracy numbers with no decomposition reads as incomplete. The methods and experiments sections are where this is decided: if a reader cannot tell which design choice produced the improvement, the manuscript is not yet ready for this journal regardless of how good the headline number looks.

Check if your Neurocomputing experiments include the ablations reviewers expect →

Missing reproducibility: no code, hyperparameters, or training protocol. Neurocomputing increasingly expects computational work to be reproducible, and the most common reproducibility gap is a methods section that omits the details needed to repeat the result. The architecture diagram is present, but the hyperparameters, training schedule, random seeds, hardware, and data preprocessing are not, and there is no code link or data availability statement that actually resolves.

Because peer review here is single-anonymized and the work is judged on whether it advances learning systems, a result that cannot be reproduced from what is written is hard for a reviewer to trust. A manuscript that treats reproducibility artifacts as optional is returned more often than authors expect, and the fix is cheap: a complete training protocol and a working code or data link before submission.

Check whether your Neurocomputing methods are reproducible from the manuscript →

Scope drift toward a knowledge-systems or pure-application venue. Neurocomputing covers neural networks and learning systems, and a recurring desk return is a manuscript whose genuine contribution is a knowledge-based reasoning system, an expert system, a feature-engineering pipeline, or a domain application with a learning component bolted on for fit.

The introduction frames the work as neural computing, but the novel result is a rule-based system, an optimization heuristic, or an application outcome, and the learning model is a component rather than the advance. Associate Editors at this journal are neural-networks specialists, and they identify quickly when the learning method is the setting rather than the subject.

A manuscript whose real contribution would be evaluated more naturally by a knowledge-systems or applications reviewer is consistently identified as a scope mismatch before review, and the better target is usually a sibling Elsevier venue.

This guide tells you what Neurocomputing editors look for; a Manusights review tells you whether YOUR paper passes that screen. A Manusights review checks the learning-systems contribution, the ablation evidence, the reproducibility artifacts, and the scope framing against the editorial bar this journal applies before peer review. Paid Manusights reviews include a 60-day money-back guarantee, and we do not train models on submitted manuscripts.

Before submitting, a Neurocomputing contribution and scope readiness check tests whether your methodological novelty, ablation evidence, and scope framing clear the editorial bar this journal applies before peer review.

This guide was built from the Neurocomputing Guide for Authors on Elsevier ScienceDirect, the Editorial Manager submission requirements, the journal's stated scope on neural networks and learning systems, and Manusights pre-submission review patterns from neural-networks manuscripts. We checked the required artifacts (data availability statement, CRediT contributions, highlights), the single-anonymized review model, and the open-access charge against Elsevier's own pages, and we cross-checked review-timing ranges against SciRev community data and Clarivate JCR 2024 journal metrics.

The failure patterns describe what we see most often when neural-networks manuscripts come through pre-submission review for this journal.

Use this page before you upload, when the official instructions cannot answer the real question: whether your learning-systems contribution, ablation evidence, reproducibility, and scope framing are already defensible. Source limitation: Elsevier updates artifact requirements, charges, and policies after this review date, so confirm administrative specifics against the journal's official pages before submission. To pressure-test the manuscript itself, run a manuscript readiness check.

• IEEE Transactions on Neural Networks and Learning Systems submission process
• Pattern Recognition submission process
• For the broader cluster, see the machine learning and AI journals overview.

Before you upload, run your manuscript through a Neurocomputing submission package check to catch the contribution, ablation, and scope issues editors filter for on first read. The check is free to run (/ai-review) and takes a single upload.