Scientifydata/raw + data/cleanCoderKeep raw data read-only; generate clean data with code.
Main compute
Python or MATLAB
Quick checks
Excel + exported CSV
Delivery
Word, or practiced LaTeX
Run one past problem with the default stack. Add a solver or domain tool only when the main environment cannot handle the model, scale, or delivery format.
Choose the main environment
The decision is whether the team can reproduce results under time pressure, not which product has the longest feature list.
Python
- Default role
- Data, statistics, optimization interfaces, batch experiments, and figures
- Choose when
- The team knows pandas, SciPy, statsmodels / scikit-learn, and environment management
- Switch or add when
- Existing MATLAB templates or engineering toolboxes dominate the work
- Competition risk
- Last-minute packages, inconsistent environments, and unordered notebooks break reproduction
MATLAB
- Default role
- Matrices, numerical solving, optimization, signals, and engineering simulation
- Choose when
- The team has practiced scripts and the required toolboxes
- Switch or add when
- Data cleaning, text, ML pipelines, or experiment automation dominates
- Competition risk
- Confirm licenses and toolbox versions before the competition
Excel
- Default role
- Inspect source tables, spot-check results, build small examples, and format tables
- Choose when
- The writer needs to verify fields, units, and a few calculations
- Switch or add when
- Repeated cleaning, parameter search, or complex formula chains appear
- Competition risk
- Hidden formulas and manual overwrites are difficult to audit
Specialist solver
- Default role
- Large linear / integer programs, constrained optimization, and feasibility evidence
- Choose when
- Basic libraries miss the time limit and the team can interpret status, bounds, and gaps
- Switch or add when
- Build a small baseline with SciPy, PuLP, OR-Tools, or MATLAB first
- Competition risk
- An optimal value without the model and solver log is not explainable
Word / LaTeX
- Default role
- Equations, figures, references, and final PDF
- Choose when
- Use the format the writer has already practiced
- Switch or add when
- Use LaTeX when long-document references matter and the team already has a template
- Competition risk
- Learning a typesetting system during the event consumes writing time
Start by problem type
Build an explainable baseline first, then upgrade for scale or validation needs.
01
Evaluation, statistics, prediction
Python + Excel + Word
pandas, SciPy, statsmodels / scikit-learn, Matplotlib
Why it works
One script can clean, fit, validate, and export figures; Excel only spot-checks.
Hand to the writer
Dictionary, clean CSV, errors, residuals, parameters, and seeds.
02
Planning and optimization
Python or MATLAB + solver
SciPy / PuLP / OR-Tools, or MATLAB Optimization Toolbox
Why it works
Confirm variables, constraints, and objective before upgrading the solver.
Hand to the writer
Formulation, input, status, gap, runtime, and solution table.
03
Differential equations and simulation
MATLAB or a practiced Python scientific stack
MATLAB / Simulink, or SciPy integrate + NumPy
Why it works
Reuse practiced solvers, tolerances, and plotting templates.
Hand to the writer
Equations, initial values, tolerances, trajectories, and sensitivity plots.
04
Graphs, routing, networks
Python + NetworkX / OR-Tools
NetworkX, OR-Tools, pandas, Matplotlib
Why it works
Keep node identifiers consistent from input through figures.
Hand to the writer
Node-edge tables, mapping, parameters, paths, and network figures.
05
Small-sample verification
Main environment + Excel double-check
Python / MATLAB results, Excel recalculation on representative rows
Why it works
Catch units, indices, percentages, and boundary errors.
Hand to the writer
Checked rows, differences, corrections, and result version.
72-hour collaboration
Modeling, coding, and writing overlap. This cadence lets all three members move in parallel.
- 1
0–3h
Read, inspect data, choose the problem, and define a minimum model.
Research constraints, profile data, and draft the paper outline.
Freeze now
Interpretation, variables, metrics, and folder structure.
- 2
3–18h
Run a baseline, write equations, and draft assumptions and data analysis.
Record decisions; export figures from code; cite result tables in prose.
Freeze now
Runnable baseline, clean data, model v1, and initial figures.
- 3
18–48h
Improve against errors and failures; run sensitivity or robustness checks.
Writer questions results while modeler checks equations and coder retains logs.
Freeze now
Claim tables, figure numbers, parameters, and rejected approaches.
- 4
48–60h
Stop expansion; finish interpretation, limitations, abstract, and references.
Rerun key scripts from a clean folder and trace every reported number.
Freeze now
Final model, data, run command, and main paper.
- 5
Final 12h
Fix only critical issues; inspect support files and the final PDF.
Check format, numbers, attachments, names, abstract, and conclusion.
Freeze now
Submission checksum, PDF, code attachment, and backup.
Retain at least these six handoff artifacts
A reliable stack lets another member locate each number and rerun it.
src/q1_*.py or .mCoderSplit by question and document dependencies and run order.
results/baseline.csvModeler + coderStore parameters, metrics, units, and model version.
results/sensitivity.csvModelerStore perturbation ranges, failures, and claim boundaries.
figures/figure_*.pngCoder + writerMatch paper numbering and retain generation scripts.
notes/decision-log.mdWriter / leadRecord decisions, rejected approaches, and open checks.
Scientify delivers a first-prize-level modeling submission end to end
As a mathematical modeling Agent, Scientify orchestrates Python, MATLAB, and other computing languages to decompose the problem, select models, write and run code, validate results, create figures, and produce the paper inside one cloud workspace.
- Turn the problem into an executable modeling plan and team workflow
- Orchestrate Python, MATLAB, and other languages to build and solve models
- Run baseline, sensitivity, and stability experiments and improve the result
- Produce figures, result explanations, and a complete competition paper
Deliver the complete model, runnable code, experiment results, figures, and a national first-prize-level paper.
The page uses recurring findings across official material, award-team reviews, and independent posts.
China Student Online
2020 CUMCM B award team review
All members model and write; tasks remain flexible and meetings leave records.
UCASS School of Economics
National first-prize experience
Writing overlaps modeling; the writer starts profiling data and figures early.
CUMCM official site
Competition experience material
Plan the paper, solve with code, and review the final format as a team.
Hou Kaifa
Roles, preparation, and resources
Master at least one of MATLAB or Python and understand adjacent roles.
Modeling software FAQ
Python or MATLAB for mathematical modeling?
Use the one your team has already used to complete a past problem. If experience is equal, choose Python for connected data and ML workflows, or MATLAB for concentrated numerical and engineering toolboxes.
Can a CUMCM team use only Python?
Yes, if the team can read data, solve, validate, and plot with it. Add another tool only when a selected problem depends on a specialist environment.
Can Excel do mathematical modeling?
Use Excel to inspect data, build small examples, spot-check results, and format tables. Move repeated cleaning, searches, and multi-round validation into code.
How many algorithm templates should we prepare?
Prepare templates you understand and have run on past problems. For each one, know its input, parameters, failure status, output, and validation method.
Enter the problem and receive a national first-prize-level submission
Scientify orchestrates multiple computing languages to complete modeling, solving, validation, figures, and the paper end to end.