Is Loess Cement Alive? A Sterile-Control Test of Microbial Contributions to Collapse

Loess is the wind-blown silt that covers much of northern China. Dry, it stands in near-vertical cliffs. Wet, it can lose most of its strength within hours, which is how railway embankments slip and tunnel roofs sag. Engineers model this collapse as pure physics and chemistry: water dissolves the calcite "bridges" that glue silt grains together and thickens the water films on clay particles. Microbiologists, meanwhile, profile loess bacteria with 16S rRNA gene sequencing, a method that reads a short, lineage-specific stretch of a gene every bacterium carries and so reveals which bacteria are present. They have done this only in topsoil, and they have never measured strength. Nobody has asked whether the bacteria living inside the structural layers, 3 to 7 m deep, change how fast that structure falls apart when it is wetted over and over.

We propose a direct experiment. Undisturbed loess specimens are split into live, sterilized, and sterilized-then-reinoculated groups, then taken through repeated wet–dry cycles. Structural loss is measured with the geotechnical structural index, and the bacterial community is tracked by 16S sequencing. We run the test on natural loess and on loess stabilized with lignosulfonate, an organic by-product of papermaking. Stabilizing loess with lignosulfonate adds exactly the kind of carbon that native bacteria might consume. If the live and sterile groups diverge, collapse models and the long-term durability of bio-based stabilizers both have a missing biological term. If they do not, the field gets a clean negative result that justifies continuing to treat loess as abiotic.

Core Claim

We hypothesize that, under repeated wetting, part of the loss of loess structure depends on living microbes. We further hypothesize that this part is small in natural loess but substantial in lignosulfonate-stabilized loess, where the added organic network is a food source. The claim can be tested because the geotechnical literature supplies a well-calibrated abiotic yardstick. Research on structural parameters of loess and its experimental determination method defines a structural parameter $m_\tau = \tau_o/\tau_{rs}$. Here $\tau_o$ is the shear strength of the undisturbed soil and $\tau_{rs}$ is the strength of the same soil remolded and saturated, which serves as the "zero structure" reference. In Weinan Q3 loess this parameter falls from 3.52 at 6% moisture to 1.93 at 36%, and the paper attributes the drop entirely to calcite dissolution and clay softening. Improvement of Structural Parameters of Loess and Its Relationship with Strength Indicators collapses moisture and dry density into a single structural state variable $P_{cs}$ that fits $m_\tau/m_{\tau o}$ with $R^2 = 0.8984$.

The new contribution is causal, not correlational. A biological contribution should appear as a live-minus-sterile gap in $m_\tau$ at matched moisture and density. That gap should be undone when the native community is added back. It should also be structural loss that $P_{cs}$ cannot account for.

What Prior Work Gives and Leaves Open

The motivating work comes from both sides. On the mechanics side, the two Weinan papers above, the exponential cohesion law $c = 77.463\times10^{-0.105w}$ ($R = 0.94$) from Reasonable Support Parameters for Loess Tunnel Under Different Water Contents, and the 20.71% moisture early-warning threshold from Study on the disaster mechanism and prevention technology of embankment slip-collapse after extreme rainfall in the loess area all predict strength from moisture and density alone. None of them includes a sterile control. So none can say whether their residual scatter contains biology.

On the stabilization side, Study on the Mechanical Properties of Loess Improved by Lignosulfonate reviews how a 1–1.5% dose builds a three-dimensional cementing network. That network raises tensile strength by up to 26.1% and dynamic modulus by 60%. The review treats lignosulfonate as inert and notes that long-term durability is poorly characterized. Our lignosulfonate arm is aimed directly at that open question.

On the microbial side, Stochastic Processes Shape Bacterial Community Diversity Patterns along Plant Niche Gradients sequences 16S from Loess Plateau rhizosphere soil. It finds that random processes, mainly drift and dispersal limitation, account for 71.1–83.5% of how the community is assembled. The response of soil eukaryotic microbial communities to afforestation in mountainous area of the Loess Plateau profiles fungi and protists by 18S sequencing under land-use change. Both studies stay in the topsoil and measure no mechanics.

The stochastic-assembly result also acts as a constraint on our design. If community composition in loess is mostly random, then 16S composition may be a weak predictor of anything mechanical. For that reason we make the sterile-versus-live contrast the primary test and treat sequence-based prediction as secondary.

For methods, we reuse structural topic modeling from Exploring thematic structure and predicted functionality of 16S rRNA amplicon data. That approach represents each sample as a mixture of co-occurring bacterial subcommunities, and it outperformed raw abundances for classifying disease. We use it as a compact feature set, and we do not rely on its functional predictions, which the authors themselves call hypotheses. We leave out non-stationary network inference of the kind built on Identifying stationary microbial interaction networks based on irregularly spaced longitudinal 16S rRNA gene sequencing data. Our sampling is too sparse for it, and it would not answer the causal question.

Research Program

We will take undisturbed block samples from exploratory wells 3–7 m deep in Q3 loess of the Weinan type. We will characterize them the way the Weinan studies did: XRD mineralogy, calcimetric carbonate content, Atterberg limits, dry density (the file reports 1.28–1.42 g/cm³ at the reference site), and a remolded-saturated reference strength for computing $m_\tau$.

Specimens are assigned to a two-by-three factorial design. The first factor is soil treatment: natural loess, or loess remolded to the same dry density with 1% lignosulfonate. The second factor is biology. The live group keeps its native community. The sterile group is gamma-irradiated. The reinoculated group is irradiated and then receives a filtered cell extract from the same horizon.

Each cell of the design is taken through 0, 5, 10, and 20 wet–dry cycles. Each wet phase brings the specimen above the 20.71% instability threshold, and each dry phase brings it back to about 10%. Specimens are incubated at a constant temperature, with at least four replicates per cell and cycle count.

At each endpoint, one set of specimens undergoes triaxial shear at 100, 200, and 400 kPa confining pressure to compute $m_\tau$, cohesion, and friction angle. Paired specimens receive unconfined compression and wet-disintegration tests. From the same specimens we also measure carbonate content, dissolved calcium in the pore water, cumulative CO₂ respiration, and 16S V4 amplicon sequencing. The sequencing is paired with quantitative PCR of the 16S gene, so that community changes can be expressed as absolute abundances rather than only as proportions.

Sequencing data are denoised to exact sequence variants and assigned taxonomy with SILVA. We then summarize them in two ways: as centered log-ratio abundances (a standard transform for proportion data) and as topic proportions from the structural topic model.

Baselines

The naive reference point is the moisture-only degradation law. We fit Chen et al.'s exponential-cohesion and linear-friction form from the Fengshouling tunnel study to our own specimens, and it predicts no dependence on cycle count or biology. Any effect of biology must first survive against this law.

The strongest existing method is the $P_{cs}$ state-variable model from Improvement of Structural Parameters of Loess. It reports $R^2 = 0.8984$ against $m_\tau/m_{\tau o}$ on Weinan loess, and it is the most complete abiotic account of structure available. We refit its functional form to the live, sterile, and reinoculated specimens pooled together. It receives the same measured moisture and density that every other model receives, and it is tuned with the same cross-validation folds, which hold out whole specimens so that no specimen appears in both training and testing. We add carbonate content and cycle count as covariates, which gives $P_{cs}$ every abiotic advantage.

Beating this baseline has a precise meaning. Adding a biology indicator, or 16S features, must reduce held-out error in $m_\tau/m_{\tau o}$. The reduction must exceed the variation seen across random fold assignments. It must also hold in the lignosulfonate arm when that arm is analyzed separately.

The key ablation is the sterilization contrast itself. To separate a genuine biological effect from damage caused by irradiation, we compare sterile and live specimens at zero cycles. We also use the reinoculated arm: a real microbial effect should reappear in reinoculated specimens, whereas an irradiation artifact would not. A second ablation replaces the native extract with an equal-biomass extract from topsoil. This tests whether the effect depends on which organisms are present or only on how many cells there are.

Why not just…?

Why not just trust the abiotic models, since they already fit so well? An $R^2$ near 0.9 leaves room for a systematic effect, and more importantly the abiotic models have never been tested against a condition that removes biology. The fit was obtained at a single time point. It cannot show whether repeated cycling in live soil degrades structure faster than moisture and density predict. If our live and sterile arms do not differ once the zero-cycle irradiation offset is removed, this alternative wins, and we will report that result as the main finding.

Why not just measure respiration and carbon chemistry instead of sequencing? This is a real threat. If a microbial effect exists, cumulative CO₂ respiration and the loss of dissolved organic carbon may predict it as well as any 16S feature does, and they cost far less. The stochastic assembly reported in the Loess Plateau rhizosphere study makes it even more likely that composition adds little. We therefore test it head-on. We compare a model of $P_{cs}$ plus respiration and chemistry against the same model plus 16S features. The topsoil-inoculum ablation adds a second check: if equal respiration from a different community produces a different loss of structure, composition matters. If 16S features add nothing, our recommendation becomes a cheap respiration assay for durability screening, and sequencing is reserved for diagnosing which organisms are responsible.

Why not just run standard accelerated durability tests on lignosulfonate loess? Wet–dry durability testing would show whether stabilized loess weakens. It cannot say why. Hydrolysis, leaching, and biodegradation of the polymer would all look alike. Only the sterile arm separates them, and the answer changes the engineering response: biocide or dosage adjustments help only if the cause is biological.

Decisive Checks

The first decisive check is the primary contrast. After 20 cycles, is the live-minus-sterile difference in $m_\tau$ larger than the zero-cycle irradiation offset plus two between-replicate standard deviations? We will pre-register this threshold, because the file contains no reported number for any live-versus-sterile loess strength comparison. There is no reported result found on Lacuna for a biological contribution to $m_\tau$. We expect a small gap or none in natural loess, and a clear gap in unconfined strength and wet-disintegration resistance in the lignosulfonate arm.

The second check is restoration. The reinoculated arm should recover at least half of the live–sterile gap. If it does not, the gap is attributed to sterilization artifacts rather than to microbes.

The third check is mechanism. Where the gap exists, we ask whether it tracks carbonate loss and dissolved calcium or instead tracks a loss of lignosulfonate, which we measure by ultraviolet absorbance of pore-water extracts. We also ask whether the taxa that grow in absolute abundance, as measured by quantitative PCR, are the same ones that dominate the subcommunity topics most correlated with strength loss.

The fourth check is predictive. Using specimen-held-out cross-validation against the refit $P_{cs}$ baseline and the respiration-augmented model, we measure whether 16S features reduce error. Together, these checks turn one question — does soil biology belong in collapse models? — into a yes, a no, or a "only for organic stabilizers" answer that an engineer could act on.