27-Year Multi-Producer European Cement Characterization Dataset

Almost every cement laboratory archive dies in a filing cabinet: thousands of routine quality-control measurements, taken every shift for decades, useful to nobody outside the plant that took them. This dataset is what happens when somebody rescues one. Researchers at Bauhaus-Universitat Weimar and the Technical University of Munich assembled 476 cement samples from 23 anonymized European producers, covering laboratory deliveries from 1997 through 2024, into a single processed table of 50 variables per sample.

The split is 311 CEM I samples against 165 other records - mostly CEM II and CEM III, with a small number of CEM IV, CEM VI and special cements. Each row binds together what the lab actually measured: declared type and strength class, full oxide chemistry, density and Blaine fineness, particle-size descriptors where measured, setting times, water demand, Le Chatelier soundness, and flexural and compressive strength at 2, 7 and 28 days. Derived engineering features ride alongside - lime saturation factor, silica and alumina moduli, equivalent alkali and corrected Bogue phase fractions - so the table supports clinker-modeling work without re-deriving anything.

The multi-producer, multi-decade design is the rare part. Single-factory archives cannot tell you whether a chemistry-to-strength relationship generalizes across plants; this one can. Get a sample of this dataset and pull the CEM I rows first - they are the backbone the accompanying machine-learning study was built on.

One row per cement sample, flat tabular, no parsing gymnastics. Three illustrative rows from the same anonymized producer:

company        : Company_1         cement_type    : CEM I
strength_class : 52.5              delivery_date  : 2001-05-28
density        : 3.132 g/cm3       blaine         : 4140 cm2/g
chem_CaO       : 0.625             chem_SiO2      : 0.205
compressive_28d: 58.7 MPa          is_cem1        : True

company        : Company_1         cement_type    : CEM I
strength_class : 52.5              delivery_date  : 2004-03-12
density        : 3.118 g/cm3       blaine         : 4230 cm2/g
chem_CaO       : 0.629             chem_SiO2      : 0.201
compressive_28d: 63.4 MPa          is_cem1        : True

company        : Company_1         cement_type    : CEM II/A-LL
strength_class : 52.5              delivery_date  : 2018-10-16
density        : 3.095 g/cm3       blaine         : 6080 cm2/g
chem_CaO       : 0.632             chem_SiO2      : 0.197
compressive_28d: 61.6 MPa          x50            : 6.01 um

Values above quote the delivered table directly. Read the three rows together and the longitudinal design shows itself: within one anonymous producer, Blaine fineness climbs from 4140 to 6080 cm2/g between a 2001 CEM I and a 2018 CEM II/A-LL while calcium oxide holds near 0.63 and 28-day strength stays in a narrow 59-63 MPa band. That is seventeen years of grinding technology and binder evolution captured in comparable units - the kind of panel a single-year snapshot simply cannot give you. Chemistry fractions are stored on a 0-1 scale, with weight-percent documentation carried in the shipped data dictionary.

Fifty variables make up the full schema; the twenty-four columns below are the verified spine covering identity, chemistry, physicality, setting behaviour and mechanical performance. Every field carries its unit, storage scale and derivation notes in the publisher's own data dictionary, which travels with the table. Columns beyond the spine - the minor oxide assays, the remaining particle-size percentiles, the rest of the Bogue set - are itemized in the final row.

Geography - Europe, sampled across 23 cement producers. Producer identities are anonymized by design and the countries involved are deliberately not disclosed, so treat this as a continental panel rather than a country-mapped one. If your question needs national attribution, pair it with asset-level sources such as the global cement and concrete tracker.

Temporal - laboratory deliveries from 1997 to 2024. Twenty-eight years is long enough to span several generations of grinding technology, clinker chemistry adjustment and the industry's gradual shift toward blended CEM II and CEM III binders, which means the table supports genuine trend work rather than a single-period correlation exercise.

Granularity - one row per cement sample: 476 rows, 50 variables, 311 of them CEM I. Sample-level resolution is what makes the strength-modelling use case work; aggregate tables average away exactly the variance a regression needs.

API, files, or your warehouse. Daily, weekly, or hourly.

• Machine-learning strength prediction - train and validate models that infer 28-day compressive strength from early-age results and routine chemistry. The publishers themselves framed the dataset as a probe of how far routine cement characterization can be pushed, which makes it a ready-made benchmark; methodology notes live on our data scientists use cases page.
• Binder evolution research - track how CEM I versus blended cements drifted across 1997-2024 in fineness, sulfate balance and setting behaviour, with producer held constant via the anonymized labels.
• Standards and specification work - examine how declared strength classes map onto measured 2, 7 and 28-day performance across many producers at once, the cross-plant view single-lab studies lack.
• Mix design and constructability studies - standard-consistency water demand, initial and final setting times and soundness arrive per sample, so rheology-relevant inputs need no separate sourcing.
• Decarbonization assessments - the 165 non-CEM-I rows give clinker-substituted binders the same measured treatment as portland cement, useful baseline material for lower-carbon binder performance claims.

Data scientists and ML engineers get a cleaned, dictionary-documented, sample-level table with derived features already computed - the tedious half of cement modelling done before you open a notebook. Materials R&D and quality engineers get a cross-producer reference distribution for every routine lab measurement, something no individual plant's archive can provide. Developers and data-product builders get a compact, stable schema that drops straight into feature stores and dashboards. Journalists, academics and students get citable laboratory evidence behind claims about European cement quality trends; more angles on our journalists academics students page.

Most of this industry's data measures tonnes, euros or plants: production series, price trackers, asset registries. This dataset measures the material itself, at laboratory resolution, per sample. The nearest neighbor is the classic 1,030-row Concrete Compressive Strength Dataset - we run the comparison in detail on our vs Concrete Compressive Strength Dataset page; the short version is real multi-producer production samples versus a smaller, coarser mix-proportion table.

Used together they bracket the problem: production and asset sources say who makes cement and how much; this table says what the product actually tests like. The full slate sits on our best construction materials datasets ranking.

Three things. First, anonymity cuts both ways: producer labels let you hold a company constant across decades, but no country or region attribute exists anywhere in the table, so geographic segmentation beyond 'Europe' is off the table.

Second, expect structural zeros rather than dirt: particle-size descriptors are populated only where PSD was measured, and the shipped data dictionary documents every unit, storage scale and missing-value convention - read it before writing imputation logic, because fraction-valued columns stored 0-1 will silently mislead anyone assuming weight percent.

Third, this is a longitudinal archive through 2024 deliveries, not a rolling series - plan analyses around its fixed historical window and let scheduled delivery handle freshness of your copy.