eleticle

Core No. 09Logged 20 Nov 2026Code & data →

A model predicts grinding throughput with R² = 0.962. What did it actually learn?

Public dataset · 5,000 rows · every value rebuilt from seed 42

We fitted a straight-line regression to a public file of 5,000 grinding records, then tested whether a random-number generator and a formula could reproduce the file, and whether a mill could have made its rows.

What we found

  • Made by a generatorAll 75,000 numbers in the file, throughput included, come out of a random-number generator started at seed 42.
  • Throughput is a formulaEleven inputs weighted and added, a sine wave in feed size and random noise, stretched to run from 170 to 350 t/h. The regression learned the formula's straight-line part, not a mill.
  • Why 0.962A straight line follows 96 % of the formula's swing. The sine wave and the noise are the rest.
  • More out than inIn 1,732 of the 5,000 rows, throughput is higher than the feed rate. Read as fresh feed and final product, both in dry tonnes per hour, no mill at steady state puts out more than it is fed.
  1. Method
  2. 01The file
  3. 02The model
  4. 03The score
  5. 04The mill
  6. 05Before you trust it

Method

A public file, rebuilt from seed 42

  • RealThe public file“Grinding Throughput in Mining Operations”, downloaded from Kaggle on 9 Nov 2026 under the Apache 2.0 licence: 5,000 rows, 15 columns, no timestamps. Its page does not say how the rows were made.
  • RebuiltSeed 42A short script starts NumPy's legacy generator at seed 42, draws the inputs and the noise, adds up the formula, stretches it to 170–350 t/h and rounds. Its output has the same SHA-256 as the download. We found the formula by fitting the file; whether its author wrote it in this form we cannot know.
  • ModelOrdinary least squaresOn the 14 inputs, with an intercept: R² 0.962 on all 5,000 rows, and 0.958 to 0.965 on five held-out folds of 1,000 rows, 0.962 on average. An effect is the learned weight times the input's range.
  • ScopeThis file onlyThis shows what a model learns from this file. It says nothing about how any real mill behaves.

The script that rebuilds the file and fits the regression needs only Python's standard library. It is in the article's public repository, under the MIT licence. The dataset stays on its Kaggle page.

01 · The file

Seed 42 rebuilds the whole file

The dataset's page describes 5,000 records of a grinding process: 14 ore and operating variables, and throughput in t/h. Every column has the shape random draws give. Start NumPy's random-number generator at seed 42, follow six steps, and out comes the same file, byte for byte.

The file's 15 columns · histograms5,000 rows each
Fifteen histograms. All fourteen inputs are flat boxes, evenly spread between two round limits, such as feed rate from 180 to 380 t/h and ore hardness from 80 to 180 MPa. Throughput, from 170 to 350 t/h, is bell-shaped.

Each input is a flat box: an even draw between two limits, the mark of a random-number generator. Plant tags bunch around the way the mill is run. Throughput is the bell that a weighted sum of random draws makes.

  1. Seed42, NumPy
  2. 14 inputseven draws between two limits
  3. Noisebell-curve draws, sd 4
  4. Throughput11 inputs weighted + 5·sin(feed size) + noise
  5. Stretchto run from 170 to 350 t/h
  6. Roundto 3 decimals, then 2

75,000 of 75,000 values match · same SHA-256 as the download

02 · The model

The regression found the formula

Fit a straight line to the 14 inputs and it scores R² = 0.962 on rows it has not seen. Each ring is what the formula does to throughput across an input's range; each dot is what the regression learned. Change the formula and watch the dots follow.

Effect on throughput · t/hR² 0.962 on unseen rows · axis ±80 t/h
For each input, the change in throughput across its range: a ring for the formula that made the file and a dot for what the regression learned. Every dot sits on its ring except feed size's: feed rate 68 t/h, mill speed 48, ore density 40, mineral grade 22, ore hardness minus 19, mill load 14, grinding media 13, water flow 11, power draw 10, vibration minus 9, moisture minus 8; temperature and operating hours none. Feed size has no straight-line weight in the formula, only a sine wave; the regression gives it 2 t/h.

In this formula, feed size has no straight-line weight, only a sine wave of ±5 t/h that repeats every 6.3 mm.

The file's own formula. On rows it never saw, the regression scores R² 0.962, and each dot sits on its ring except feed size's, which catches part of the sine wave.

It learned what it was given

Mineral grade adds 22 t/h across its range because the formula says so, and temperature adds nothing for the same reason. Nothing in the file can tell the model whether either holds on a mill.

What a straight line misses

A straight line cannot follow the sine wave in feed size, and nobody can predict the noise. Together they are the 3.8 % of the swing the model gets wrong.

03 · The score

Where 0.962 comes from

R² is the share of throughput's swing a model reproduces. Here the swing has three parts, and each kind of model stops at a different one.

Throughput's swing, in partsR² 0.962 · 0.978 · 1.000 on all rows
Throughput's variance in three parts: 96.1 % from the straight-line part of the formula, 1.7 % from the sine wave in feed size and 2.2 % from the noise. Zoomed on the last 5 %: the regression reaches R² 0.962, a model that also knew the sine wave would reach 0.978, and replaying seed 42 reaches 1.000, as would a lookup table of these rows; on rows not yet drawn only the replay would.

So 0.962 measures how much wave and noise went into the formula, not how well the model knows a mill. Reverse the hardness effect in 02 and the score moves by 0.001: R² cannot tell whether a formula is physical.

04 · The mill

No mill could have made these rows

Read the two columns as fresh feed and final product, in dry tonnes: at steady state a grinding circuit's product equals its feed. Every row above the line reports more product than feed.

Throughput against feed rate · t/h1,732 of 5,000 above the line
A cloud of 5,000 points with a shallow upward slope. 1,732 of them lie above the line where throughput equals feed rate. In the formula, each extra t/h of feed adds 0.34 t/h of throughput.

In the formula, each extra t/h of feed adds 0.34 t/h of throughput. At steady state, product follows fresh feed one for one until the circuit reaches a limit. Only internal streams, carrying the circulating load, run above it.

Inputs that never move together

On a mill, water is usually added in ratio to the feed to hold the slurry density, and power draw follows the charge and the speed. In this file the strongest link between any two inputs is a correlation of 0.03.

No time, so no later

The rows have no timestamps. Nothing can be tested on a later month, after a liner change or on a new ore, because the file has no later.

05 · Before you trust it

What to see before a model goes near a mill

Ask for these before the score. This file has none of them.

  1. Where each number came fromThe tag, the instrument, the units, the sampling and any filtering.This file: a random-number generator
  2. A test on later dataTrain on earlier months and score on later ones, using only inputs known when the prediction is made. Report the error in t/h, against the tolerance you need.This file: no timestamps
  3. Throughput at a stated grindTonnes per hour mean little without the product size they were ground to. A coarser grind can allow more tonnes when grinding energy is the limit.This file: no product size
  4. Balances that closeThroughput against feed, kWh per tonne against the mill's power, water against slurry density.This file: 1,732 rows with more throughput than feed
  5. Inputs that move like a millLinks that match the control loops and the physics: water with feed, power with load.This file: none stronger than 0.03
  6. Effects the mill confirmsThe model's answer to more speed or more feed matches a step test or the operators' experience.This file: nothing to check against

R² is the last check, not the first. A model that passes these six makes its score worth examining.