We grew maize (zea mays, Reid's Yellow Dent Open Pollinated Corn Seed, Bradley Seed Brand) in a Yale Science Building research greenhouse (affiliated with Yale University's Marsh Botanical Garden's plant growth facility) controlled by an Argus Prime automation system with and without basalt amendments under average growing season conditions. We measured greenhouse gas fluxes, soil pH, topsoil Sikora buffer pH, pore water alkalinity, temperature, soil moisture, and total nutrient of the soil and corn plant matter. We performed two iterations of the experiment (Run 1, Run 2) in the same mesocosms. System settings and measurements taken for each iteration of the experiment are listed in Table 1.

All maize was grown in 121-liter containers with a 55.88 cm diameter. We chose to use large containers relative to those used in other ERW mesocosm experiments [e.g., ~55-liter (Amann et al., 2020), and ~7.8-liter (Buckingham et al., 2022)] given that small column experiments may lead to differences in soil and system behavior that can alter ERW rates and may alter N₂O fluxes. In each container, we planted 8 corn seeds at close to standard agricultural densities (8–12 inches apart) and thinned to 4 stalks after 2 weeks if necessary, selecting the smaller stalk to be removed in all cases. We applied fertilizer at a rate of 242 kg N/ha both to ensure a visible N₂O flux peak (as N₂O fluxes increase with increased fertilizer applications rates Huang et al., 2014; Roy et al., 2014) and because many farms in major corn producing states apply high rates of nitrogen fertilizer (Xia et al., 2021). Each container received fertilizer via injection of urea-ammonium-nitrate fertilizer in a row 5 cm away from the seeds to minimize risk of seed burning.

We also measured or modulated other factors in addition to soil pH that have been documented to impact N₂O fluxes, including soil moisture, soil texture, crop type, fertilizer timing and application rate, soil organic carbon content, temperature, and upper soil modification (tilling practice) (Dobbie and Smith, 2001; Stehfest and Bouwman, 2006; Shcherbak et al., 2014; Jiang et al., 2022; Vangeli et al., 2022). Soil moisture is typically observed to have a positive correlation with N₂O fluxes, as it decreases the soil oxygen concentration and stimulates denitrifying activity (Dobbie and Smith, 2001; Butterbach-Bahl et al., 2013). Soil organic carbon content also plays a significant and complex role in soil nitrogen cycling but is typically positively correlated with N₂O fluxes (Stehfest and Bouwman, 2006). Soil texture and clay content impacts N₂O fluxes as more porous soils have increased oxygen availability, stifling denitrification and decreasing N₂O fluxes (Mei et al., 2018). Tilling, on average, is likely to increase soil oxygen availability and decrease soil organic carbon content, which together should lead to decreased N₂O fluxes (Mei et al., 2018). Fertilizer application rate also directly impacts the amount of bioavailable nitrogen in the soil that can be emitted as N₂O (Roy et al., 2014; Shcherbak et al., 2014), and timing fertilizer application during the growing season decreases N₂O emissions as plant uptake reduces the potential of N₂O formation from soil nitrogen pools (McGill et al., 2018; Shakoor et al., 2021). Temperature has also been found to be positively correlated with N₂O flux, as it stimulates microbial community activity and soil respiration (Dobbie and Smith, 2001; Lai et al., 2019).

We kept variables affecting N₂O fluxes as constant as possible with the Argus Prime automation system between containers within each iteration of the experiment except for soil pH, which was adjusted via the basalt amendments. The system was programed to mimic average July conditions in Spring Grove, Illinois, USA. We chose to aim for 60% saturated soil moisture—in our mesocosms corresponding to a volumetric water content (VWC) of 25—in order to ensure a significant N₂O flux (Sehy et al., 2003; Skiba and Ball, 2006; Roy et al., 2014). We maintained consistent soil moisture via automated irrigation lines dispensing reverse osmosis water through a Netafim NetBow 10-in drip ring via the Argus Prime system, and a porous drainage medium placed at the bottom of each container. We ended the experiment once the N₂O fluxes returned to baseline values, typically 24 to 29 days after fertilizer addition.

Basalt feedstock was sourced from waste fines at the East Haven Trap Rock Quarry and prepared by passage through a 0.79 mm sieve. A summary of the following analyses can be found in Table 2. We performed particle size distribution analysis on the sieved basalt with a Microtrac Flowsync Particle Size and Shape Analyzer (Supplementary material 1). We performed BET analysis for the sieved basalt on an Anton Paar Nova 800 (Supplementary material 2). Major and trace element analysis of sieved basalt feedstock was performed using the Ultratrace Aqua Regia complete chemical method at Actlabs (41 Bittern Street, Ancaster, Ontario, L9G 4V5 Canada) (Supplementary material 3). Based on petrographic work, the feedstock is a weakly carbonized metabasalt with traces of secondary quartz veins. Applied feedstock was homogenized in the upper 10 cm of soil in order to simulate disc tillage. Control containers received the same “tilling practice” without any basalt application to surface soil.

Weekly pore water samples were taken using Rhizon samplers inserted at three different soil depths (15 cm, 35 cm, and 50 cm) for pore water alkalinity measurements. We calculated alkalinity using a Thermo Scientific Orion Star T920 redox titrator with 0.0501N HCl as the titrant. We performed reproducibility tests and found an error of 1.4% between samples of our sample size (4 mL). Weekly soil moisture (in the form of VWC) measurements were taken at three different soil depths (15 cm, 35 cm, and 50 cm) using a Spectrum Technologies TDR 150 soil moisture meter with an accuracy of ±3.0% VWC. Surface soil was also collected to measure soil pH (with a 1:2.5 soil: water ratio) and to measure Sikora buffer pH. The buffer pH was measured via the Sikora method (Sikora, 2006), with reproducibility tests showing a standard deviation of ±0.024 pH units. The surface soil temperature of each container was continuously monitored by the Eosense automated soil flux chambers (eosAS-LT/LO). All total nutrient analyses of the soil and corn plant matter were performed by Agvise Laboratories, Inc. (804 Highway 15 W, P.O. Box 510, Northwood, ND 58267) (Supplementary material 4, 5).

N₂O fluxes were measured with a Picarro Cavity Ringdown Spectrometer model G2508 paired with 10 Eosense automated soil flux chambers (eosAS-LT/LO) and an Eosense recirculating multiplexer (eosMX) (Anthony and Silver, 2020; Eosense, 2020, 2022). A 10-min measurement period was used, with a linear fit and a pre- and post-delay of 1.5 min, to ensure high-resolution flux measurements while measuring the maximum number of samples per container each day. The mass spectrometer was calibrated at the start of the iterations.

We performed two iterations of the experiment (Run 1, Run 2) with the same conditions as described above. After Run 1, the corn plants were terminated at the end of the peak N₂O flux period by cutting the stalk at ground level, to prepare the experiment for Run 2. The second iteration (Run 2) only received new corn seeds and fertilizer, and it did not receive new soil, tilling, basalt, leaf compost, or acid. Run 2 can be treated as a second growing season on soil that had previously been amended during Run 1.

To calculate total emissions, the flux measurements for each container were integrated over the length of the run using the trapz function from scipy.integrate in Python. Difference in means of cumulative emissions was tested for statistical significance with the two-tailed t-test function from scipy.stats in Python. Manual measurements were linearly interpolated to line up with the continuous flux measurements. The instantaneous flux data, weekly manual measurement data (excluding alkalinity), and application rate data were then fed into the machine learning algorithm in R to predict the instantaneous N₂O fluxes from various levers and to assess the relative importance of those levers in regulating the N₂O fluxes.

Given the high-dimensional nonlinear pattern in our data, we adopted a machine learning technique to predict N₂O fluxes and to quantitatively evaluate the relative importance of each parameter in driving variation in N₂O flux. Compared with traditional numerical methods, machine learning has the advantage of being able to approximate arbitrary nonlinear functions with large sets of intertwined parameters, in situations where a physio-chemical based algorithm might be insufficient. Here, we employ the random forest (RF) algorithm (Ho, 1995; Breiman, 2001), a decision tree-based algorithm in which an ensemble prediction is produced from many sub-models (i.e., decision tree) with each “tree” model making independent predictions for the variable of interest (in our case, N₂O flux) using assigned input variables (i.e., predictor variables, such as basalt application rate, temperature, soil pH, etc.). The construction of the RF model is conducted in R (R Core Team, 2017) using the “ranger” package (Wright and Ziegler, 2017). We first randomly divide the data into a training dataset (75% of the entire dataset) and a test dataset (25% of the dataset). The training dataset is then used to build our RF model and the test dataset is used to test the performance of our trained RF model on unseen data. During the training process, we form an ensemble of 600 trees with a minimal node size of 3. To determine the relative importance of each feature in driving the change of predicted N₂O flux, we adopt the permutation method embedded within the RF model by first calculating the prediction accuracy using the original data then calculating the prediction accuracy again using perturbed data (sequentially perturbing a single predictor variable at a time). The difference between the two accuracy values is the permutation importance for this specific predictor variable and the average difference among all trees gives the RF permutation importance of this variable.

The N₂O fluxes in Run 1 returned to baseline values 24 days after beginning this iteration of the experiment (Figure 1). The N₂O fluxes in Run 1 (between 1.5 and 2.25 nmol N₂O/m²s) were similar to common agricultural N₂O fluxes during the N₂O spike after nitrogen fertilizer application (Lu et al., 2021). There was an obvious negative correlation between N₂O flux and soil pH—in basalt-amended mesocosms, soil pH increased while N₂O fluxes were smaller (Figure 1A).

The cumulative N₂O emissions were higher in the control group than the basalt amended containers (Figure 2 and Table 3) and followed the negative correlation between N₂O emissions and soil pH (Figure 3A). There was also an inverse correlation between porewater alkalinity and N₂O emissions (Figure 3D). In Run 1, there weren't any obvious correlations between N2O emissions and soil moisture or temperature (Figures 3B, C). The statistical analysis revealed that Run 1 did have statistically significant difference of means between the basalt amended containers and the control containers, and it showed that Run 1′s results rejected the null hypothesis (Table 4).

In Run 1, soil pH increased in the basalt amended containers and remained consistent in the control containers (Supplementary Figure 1 and Supplementary material 6). Buffer pH was higher in the basalt amended containers than in the control containers (Supplementary material 6). The alkalinity at 15 cm was variable in the first 14 days of the experiment but then began to increase slightly in the control containers and more rapidly in the basalt amended containers (Supplementary Figure 2 and Supplementary material 6). Temperature in Run 1 was consistent for each container diurnally but varied spatially with the largest disparity between the containers being ~3°C (Supplementary Figure 5 and Supplementary material 7). The soil moisture was consistent through time for each individual container but varied between containers with the range being ~17 VWC (Supplementary Figure 6 and Supplementary material 6).

We found statistically significant differences in mean value between the control and basalt amended containers in the following categories: pH, Sikora buffer pH, ammonium-nitrate, calcium, sulfur, zinc, iron, manganese, copper, base saturation percent calcium, exchangeable acidity, and calcium carbonate equivalency (Supplementary material 8). Of these, relative to the control containers, the basalt amended containers had higher pH, Sikora buffer pH, calcium, sulfur, base saturation percent calcium, and calcium carbonate equivalency. Of the statistically significant differences, relative to the control containers, the basalt amended containers had lower ammonium-nitrate, zinc, iron, manganese, copper, and exchangeable acidity.

We also found a statistically significant difference in mean between the initial and final measurements on the basalt amended containers in the following categories: pH, Sikora buffer pH, salts, nitrate-nitrogen, calcium, sulfur, iron, manganese, copper, chlorine, cation exchange capacity, base saturation of potassium and calcium, exchangeable acidity, and calcium carbonate equivalency (Supplementary material 8). Of these, there were increases in pH, Sikora buffer pH, salts, nitrate-nitrogen, calcium, sulfur, chlorine, cation exchange capacity, base saturation percent calcium, and calcium carbonate equivalency. There were statistically significant decreases in iron, manganese, copper, potassium base saturation, and exchangeable acidity.

Between the initial and final measurements on the control containers, we found a statistically significant difference in mean across the following categories: organic matter, nitrate-nitrogen, potassium, sulfur, chlorine, and potassium base saturation (Supplementary material 8). Of these, there were increases in organic matter, nitrate-nitrogen, sulfur, and chlorine. There were statistically significant decreases in potassium and potassium base saturation.

We found no statistically significant difference in any plant tissue nutrient values between the control and the basalt amended containers (Supplementary material 8).

The N₂O fluxes returned to baseline values 29 days after beginning this iteration of the experiment (Figure 4). Control 1 (container 2) had an anomalously high N₂O flux (between 4 and 14 nmol N₂O/m²s) at the beginning of the experiment. All other containers had expected values for N₂O fluxes ranging from 0.25 to 2.5 nmol N₂O/m²s. The N₂O emissions for Run 2 did not strongly correlate with soil pH or soil amendment, but they did correlate with soil moisture (Figure 5C). To better visualize the data in case control 1′s N₂O flux was anomalous, we show the N₂O fluxes and cumulative N₂O emissions without control 1 for Run 2 (Figures 4, 5). Plots including control 1 are located in the supplemental info (Supplementary Figures 9, 10). The average cumulative N₂O emissions from the basalt amended containers were lower than those from the control containers (Figure 6 and Table 3). However, there was no statistically significant difference between the means of the N₂O emissions from the basalt amended containers and the control containers (Table 4).

Soil pH and buffer pH both remained relatively constant through time (as no new basalt was added) and were both higher in the basalt amended containers than the control containers (Supplementary Figure 11 and Supplementary material 9). Soil moisture for each container was relatively stable, but between containers the largest range was ~20 VWC (Supplementary Figure 12 and Supplementary material 9). The temperature in Run 2 behaved similarly to Run 1 and was consistent for each container diurnally but varied spatially with the largest disparity between the containers being ~3°C (Supplementary Figure 15 and Supplementary material 10).

The RF framework yielded an R² of 0.97 on the training data and an R² of 0.82 on the test data (Supplementary Figures 16, 17). The permutation importance technique performed on the data from Run 1 indicated that Sikora buffer pH (negative correlation) had the highest relative importance, followed by basalt application rate (negative correlation), soil pH (negative correlation), time (positive correlation), middle soil moisture (negative correlation), top soil moisture (positive correlation), bottom soil moisture (negative correlation), then temperature (negative correlation) (Figure 7A and Supplementary Figure 20).

For Run 2, the RF framework yielded an R² of 0.99 on the training data and an R² of 0.94 on the test data (Supplementary Figures 21, 22). In Run 2, the permutation importance technique performed on the data indicated that top soil moisture (positive correlation) had the highest relative importance, followed by middle soil moisture (positive correlation), Sikora buffer pH (negative correlation), soil pH (negative correlation), bottom soil moisture (positive correlation), time (negative correlation), basalt application rate (negative correlation), then temperature (negative correlation) (Figure 7B and Supplementary Figure 25).

Despite the basalt application rate being a binary variable (0 or 5 tons basalt per acre) within our data set, we chose to include it when training and testing the data with the machine learning framework to collate basalt-related impacts that go beyond pH or other weekly measured responses (e.g., soil structure, nutrient release, etc.). Analyses performed excluding basalt application rate as a variable are included in the supplemental info (Supplementary Figures 26–37).

Because Runs 1 and 2 were similar and Run 2 can function as another growing season on the same soil as Run 1, we analyzed the combined data with our machine learning framework. It yielded an R² value of 0.98 for the training data and an R² value of 0.91 for the test data (Supplementary Figures 38, 39). In the combined analysis, the permutation importance technique performed on the data indicated that Sikora buffer pH (negative correlation) had the highest relative importance, followed by middle soil moisture (positive correlation), basalt application rate (negative correlation), soil pH (negative correlation), top soil moisture (positive correlation), time (positive correlation), bottom soil moisture (negative correlation), then temperature (negative correlation) (Figure 8 and Supplementary Figure 42). Analyses performed excluding basalt application rate as a variable are included in the supplemental info (Supplementary Figures 43–48).

Run 1 had a N₂O peak period (24 days) consistent with those seen in agricultural settings (Roy et al., 2014; Maier et al., 2022). The N₂O fluxes (between 1.5 and 2.25 nmol N₂O/m²s) were also similar to typical agricultural values (Lu et al., 2021). The fact that soil pH, buffer pH, and pore water alkalinity all increased through time are all signals for significant basalt weathering. As expected, there was also an obvious negative correlation between soil pH and N₂O flux, likely due to the links between soil pH and microbial activity and nitrous oxide reductase assembly (Stevens et al., 1998; Bergaust et al., 2010; Liu et al., 2010; Bakken et al., 2012; Barton et al., 2013a,b; Samad et al., 2016; Abalos et al., 2020; Vázquez et al., 2020). Since alkalinity is released as basalt weathers, and if basalt application increases soil pH and decreases N₂O fluxes, then we would naturally expect an alkalinity increase with decreased N₂O fluxes.

The RF framework was successful in predicting N₂O fluxes, with an R² of 0.82 on the test data, suggesting that the current parameters fed into the machine learning framework can largely capture the variation of observed N₂O flux. Because Sikora buffer pH, basalt application rate, and soil pH were ranked as the top three in relative importance for predicting N₂O fluxes, this indicates that basalt application played a key role in mitigating N₂O fluxes during this iteration of the experiment. The middle soil moisture was ranked as more important for predicting N₂O fluxes than the top soil moisture. While top soil moisture was positively correlated with N₂O fluxes, as expected, middle soil moisture was unexpectedly negatively correlated with N₂O fluxes. Time showed a nonlinear relationship with N₂O fluxes that Spearman's rank coefficient was unable to capture, but the analysis indicated that time was positively correlated with N₂O fluxes. The fact that soil pH had a slightly smaller correlation coefficient than basalt application rate could indicate that basalt application exerts important influences on N₂O production beyond increasing the pH, for instance changing soil structure in ways that reduce N₂O production.

Run 2 also had a N₂O peak period (29 days) that was consistent with average agricultural N₂O peaks after fertilization. Additionally, all of Run 2′s containers had N₂O fluxes (between 0.25 and 14 nmol N₂O/m²s) within average expected agricultural ranges (1–50 nmol N₂O/m²s) (Huang et al., 2014; Huddell et al., 2021; Su et al., 2021; Maier et al., 2022). These fluxes were slightly higher than N₂O fluxes from Run 1, which we attribute to soil settling between iterations of the experiment given that we did not homogenize the top 10 cm of soil during this iteration of the experiment. With a more compact soil structure, there would be less aeration in the soil and therefore enhanced denitrifier activity, leading to higher N₂O fluxes. This may also explain why the soil moisture was more consistent within each container relative to Run 1. The cause of the higher N₂O fluxes in control 1 (4–14 nmol N₂O/m²s) relative to the other containers (0.25–2.5 nmol N₂O/m²s) is not clear, as it did not have a notably lower soil pH, higher temperature, or higher soil moisture than any of the other control (or basalt amended) containers. However, we have poor constraints on soil structure in the mesocosm, and this can significantly affect the extent of soil anoxic conditions and in particular the development and maintenance of anoxic microzones that could promote N₂O production.

For Run 2, the RF framework was successful in predicting N₂O fluxes again, with an R² of 0.94 on the test data. However, in this iteration of the experiment, top and middle soil moisture (which were both positively correlated with N₂O flux) were ranked higher than Sikora buffer pH, soil pH, and basalt application rate. The ability for moisture to saturate the soil more readily (from the settling of the soil) likely allowed soil moisture to be the dominant lever in Run 2. Because soil moisture was a stronger lever than Sikora buffer pH, soil pH, and basalt application rate, even though the latter three were still all negatively correlated with N₂O flux, this may explain why Run 2 did not have any statistically significant difference in mean cumulative N₂O emissions between the basalt amended containers and the control containers despite the lower cumulative N₂O emissions on average in the basalt amended containers. The lower average emissions suggest that basalt amendments decrease N₂O emissions by increasing soil pH, and the lack of statistical significance could be due to another more dominant lever—in this case, soil moisture, as suggested by the machine learning framework. This implication that soil structure plays an important role in N₂O fluxes agrees with observations suggesting that tilling practices strongly influence N₂O fluxes (Mei et al., 2018; Kim et al., 2021; Jiang et al., 2022).

The combined data from Runs 1 and 2 were also successfully predicted by the RF framework, with an R² of 0.91 for the test data. In this test, the permutation importance technique showed that while Sikora buffer pH was ranked highest in relative importance, followed by middle soil moisture, the N₂O flux was similarly sensitive to Sikora buffer pH, middle soil moisture, basalt application rate, soil pH, and top soil moisture. This suggests that both soil moisture and basalt application had strong predictive power for determining N₂O fluxes. This also suggests that ERW is a more dominant lever overall, and it decreases N₂O fluxes by removing acidity from the soil system.