Ammonia Nitrogen Removal and MAP Crystal Morphology Affected by Reaction Conditions in High-Concentration Wastewater

Abstract

The MAP (magnesium ammonium phosphate) method is a convenient and efficient approach for the recovery of ammonia nitrogen from high-concentration wastewater, with the resulting product being suitable for use as a slow-release fertilizer. The crystal morphology of MAP is a key indicator of its appropriateness for this application, yet there is a lack of systematic research on this topic. This paper explores the relationship between the efficiency of ammonia nitrogen removal, morphological characteristics of the product, and reaction conditions (i.e., pH, reaction temperature and time, phosphorus–nitrogen (n(P):n(N)) and magnesium–nitrogen mole ratios (n(Mg):n(N)), and stirring speed). The results show that the influence of the reaction parameters on the nitrogen removal efficiency decreases in this order: the pH > n(Mg):n(N) > the stirring speed > n(P):n(N). The highest ammonia nitrogen removal efficiency (97.97%) was achieved under the following optimal conditions: pH 9.5, n(Mg):n(N) = 1.3, n(P):n(N) = 1.0, a stirring speed of 150 rpm, a reaction time of 30 min, and a temperature of 30 °C. The obtained products were MAP crystals with different morphologies, which gradually transitioned from X- to needle-shaped with a decreasing crystal size as the values of the pH, n(Mg):n(N), stirring speed, and reaction time increased. These findings are relevant for both the effective removal of ammonia nitrogen from high-concentration wastewater and the control of MAP crystal morphology.

1. Introduction

Ammonia nitrogen (NH₃–N) is a critical pollutant in wastewater and is strictly regulated due to its ecological and health impacts. In China, although nutrient discharges have declined since 2016, legacy nitrogen contamination in rivers and groundwater remains widespread, often exceeding regulatory thresholds [1]. The implementation of Special Discharge Limits (SDL) has further tightened NH₃–N standards, yet compliance frequently demands energy- and chemical-intensive upgrades, posing economic and operational challenges [2]. Similar issues are reported globally, underscoring the urgent need for more efficient and resource-oriented treatment technologies.

Ammonia nitrogen wastewater is commonly classified as high-concentration, medium-concentration, or low-concentration [3]. High-concentration ammonia nitrogen wastewater (HANW) generally refers to wastewater with an ammonia nitrogen concentration of more than 500 mg/L [4], whose sources include urban domestic sewage, landfill leachate, and industrial wastewater produced by the petroleum, printing and dyeing, pharmaceutical, leather, and other industries [5]. A common and significant challenge in the efficient treatment and recovery of high-concentration ammonia nitrogen wastewater is its high biotoxicity and poor biodegradability.

The common treatment methods and applicable scenarios for ammonia nitrogen wastewater are listed in

Table A1. Comparison of treatment methods for ammoniacal nitrogen wastewater.

Method	Applicable NH4+-N Conc. (mg/L)	Removal Efficiency (%)	Cost-Effectiveness	Advantages	Disadvantages	References
Air Stripping	500–5000	80–95%	Moderate to High	Simple process, easy operation Effective for high concentrations Relatively low equipment investment	Requires pH adjustment to 10.5–11.5 Prone to scaling and clogging Poor performance in winter Cause secondary air pollution	[66,67]
Breakpoint Chlorination	50–1000	85–99%	Low to Moderate	Stable and reliable treatment effect Short reaction time Unaffected by temperature	High consumption of chlorinating agents High operating costs Generates by-products like chloramines Uneconomical for high concentrations	[68]
Ion Exchange	20–500	90–98%	Moderate	High-quality effluent Simple operation Allows for ammonia nitrogen recovery	Resin is susceptible to fouling/poisoning Requires frequent regeneration Limited application scope High pretreatment requirements	[69,70]
Struvite Precipitation (MAP)	200–8000	85–95%	Moderate to High	Recovers nitrogen and phosphorus resources Low sludge production Effective for high concentrations By-product can be used as fertilizer	Requires addition of Mg2+ and PO43− Strict pH control is necessary Potential for scaling issues Higher investment costs	[71,72]
Biological Nitrification-Denitrification	20–200	85–95%	High	Low operating costs Environmentally friendly Simultaneously remove COD Mature and established technology	Long start-up time Sensitive to high ammonia concentrations Requires an external carbon source Large footprint	[73,74]
Electrochemical Oxidation	100–2000	70–90%	Low	Mild reaction conditions No chemical reagents needed Small footprint High degree of automation	High electricity consumption Electrode passivation is common High investment costs Short electrode lifespan	[75]
Membrane Separation	50–1000	90–99%	Low to Moderate	Excellent separation efficiency Small footprint Stable effluent quality Can achieve ammonia recovery	Severe membrane fouling High operating costs Concentrate stream requires further treatment High pretreatment requirements	[76]
Catalytic Oxidation	100–5000	80–95%	Moderate	Fast reaction speed High treatment efficiency Strong adaptability Can handle high concentrations	Catalyst is prone to deactivation Relatively high operating costs Complex process Pretreatment is often required	[77,78]

. HANW represents a valuable “nutrient + energy + water” (NEW) resource [6,7]. Achieving the resource-oriented recovery of its constituent elements and its reuse as a water resource byproduct is of considerable significance [8,9]. Currently, the main resource recovery methods used include air-stripping, alkalization evaporation treatment, and magnesium ammonium phosphate (MAP, struvite) crystallization using ammonia nitrogen as a nutrient [10,11]. The air-stripping method can achieve relatively good removal in wastewater with an ammonia nitrogen concentration of 2000–4000 mg/L. The stripping efficiency is affected by a variety of factors which are difficult to control, especially the temperature [12], and the ammonia steam distillation method is generally suitable for use with wastewater with a medium ammonia nitrogen concentration of 3000–100,000 mg/L [13]. In alkalization evaporation treatment, high ammonia nitrogen concentrations may cause equipment corrosion or scaling, impairing operational stability, while low concentrations increase operating costs due to the high steam demand for heating [14]. The MAP crystallization method is suitable for medium- to high-concentration ammonia nitrogen wastewater (500–3000 mg/L) [15,16]. Struvite has the chemical formula MgNH₄PO₄·6H₂O, is an inorganic compound formed from a chemical equilibrium reaction between ammonia, phosphate, and magnesium ions [17]. MAP precipitation has been demonstrated to occur safely and stably in complex wastewater matrices [18,19]. This process can be expressed as Equation (1):

$$\mathrm{N}{\mathrm{H}}_4^{+}+\mathrm{M}{\mathrm{g}}^{2+}+{\mathrm{H}}_n\mathrm{P}{\mathrm{O}}_4^{3-n}+6{\mathrm{H}}_2\mathrm{O}\to \mathrm{M}\mathrm{g}\mathrm{N}{\mathrm{H}}_4\mathrm{P}{\mathrm{O}}_4·6{\mathrm{H}}_2\mathrm{O}+n{\mathrm{H}}^{+}, n=\mathrm{0,1},2$$

(1)

Previous studies confirm that MAP precipitation can be safely achieved in complex wastewater and that struvite is a marketable slow-release fertilizer [17,19,20]. However, research efforts to date have largely focused on two directions: (i) the treatment and nutrient recovery from complex industrial wastewaters, where issues such as ion interference (Ca²⁺, heavy metals, organic matter) and product variability are critical [16]; and (ii) pilot- or full-scale applications, emphasizing nitrogen removal efficiency, energy consumption, and life-cycle assessment of process sustainability. While these are important for practical deployment, they often bypass systematic laboratory investigations of how fundamental reaction parameters control the morphology, size distribution, crystallinity, and purity of struvite crystals.

Crystal morphology and purity strongly determine the stability, settling, and reuse value of struvite products. By contrast, phosphorus recovery studies have shown that pH and supersaturation directly influence both removal efficiency and the particle characteristics of MAP [21,22]. For example, Shaddel used lab-scale batch experiments to vary supersaturation (via pH, Mg and NH₄ concentrations) and quantified how these changes affect crystal size and morphology while also assessing phosphorus recovery and process performance [23]. In another study using digestate from anaerobic digestion, the effects of n(Mg):n(N), n(P):n(N), pH, and temperature on struvite yield and SEM-observed crystal morphology were systematically evaluated via Taguchi experimental design [24]. Similarly, studies in (synthetic or real) urine have shown that particle size, crystal habit, and separation behavior depend very sensitively on pH, supersaturation, and mixing intensity [25]. Extending such knowledge to nitrogen recovery is essential to provide the mechanistic understanding needed for optimizing product quality and supporting industrial-scale application.

The formation of struvite can include nucleation and crystallization processes, and its morphology is controlled by physicochemical factors [26]. In this study, single-factor experiments were conducted to investigate the influence of six key factors—the pH, temperature, reaction time, stirring speed, n(Mg):n(N), and n(P):n(N)—on the removal of nitrogen from high-ammonium nitrogen wastewater (HANW) via magnesium ammonium phosphate (MAP) precipitation. These experiments enabled a detailed analysis of how each factor affects the removal efficiency and product formation through crystallization during the MAP recycling process.

To better evaluate the complex relationships between the operational parameters and nitrogen removal efficiency, a comparative modeling framework was established using six machine learning regression algorithms, including linear regression, polynomial regression, Ridge Regression, Lasso Regression, Random Forest Regression, and Gradient Boosting Regression. Random Forest Regression (RFR) constructs multiple decision trees from bootstrap-resampled datasets and averages their predictions, offering high prediction accuracy, strong generalization, and effective mitigation of overfitting. This makes RFR particularly suitable for datasets with multiple variables and limited samples. Additionally, RFR’s ability to assess the feature importance allows for a clear understanding of each influencing factor’s relative significance.

Overall, this work combines single-factor experiments with machine learning analysis to provide a comprehensive understanding of the HANW nitrogen removal process using the MAP method. It helps establish a technical foundation for the efficient recovery of high-concentration ammonia nitrogen wastewater.

2. Materials and Methods

2.1. Materials

To increase the applicability of this experiment, high-concentration ammonia nitrogen wastewater was created with synthetic wastewater. Ammonium chloride (NH₄Cl), disodium hydrogen phosphate (Na₂HPO₄), magnesium chloride (MgCl₂), sodium hydroxide (NaOH), analytical-grade reagents, were purchased from Tianjin Damao Reagent Co., Ltd., Tianjin, China, detailed information is provided in

Table A2. Chemicals and reagents.

Chemical Formula	Mass	Volume	Concentration
NH4Cl	1.07 g	500 mL	560 mg/L
MgCl2	Based on experimental conditions	-	-
Na2HPO4	Based on experimental conditions	-	-
NaOH	-	-	5.0 M
KNaC4H4O6·4H2O	50 g	100 mL	500 g/L

Note: All chemical reagents used were of analytical grade and purchased from Tianjin Damao Reagent Co., Ltd., China.

. The experiments were conducted using deionized water.

2.2. Experimental Procedures

After adjusting the molar ratios of the Mg and P solutions to different values using MgCl₂ and Na₂HPO₄, respectively, 25 mL of each solution was simultaneously added to a 250 mL beaker containing 100 mL of synthetic wastewater. At the same time, the magnetic stirrer was turned on at the set speed. Subsequently, the pH was adjusted to the target value using NaOH (5.0 M) monitored continuously using a pH meter (PHSJ-3F, Rex, PR, Shanghai, China). After reaching the target time, the stirrer automatically turned off. Finally, a vacuum suction filter (SHZ-D (III), PR, Shanghai, China) was used to separate the precipitate from the reacted solution, which was dried at room temperature for 24 h.

2.3. Experimental Design

2.3.1. Single-Factor Experimental Design

According to Equation (1), the stoichiometric molar ratio of Mg²⁺:NH_4⁺:PO_4³⁻ is 1:1:1, and the reaction is highly pH-sensitive. The formation of the product is influenced by multiple factors, including the solution pH, concentrations of Mg²⁺, NH_4⁺-N, and HPO_4ⁿ⁻³ (n = 0, 1, 2), temperature, and stirring speed [27].

Studies have indicated that phosphorus predominantly exists as HPO_4²⁻ rather than PO_4³⁻ when the pH is below 8, thereby impeding the formation of MgNH₄PO₄. In strongly alkaline conditions (pH > 11), phosphorus tends to form highly insoluble Mg₃(PO₄)₂, while Mg(OH)₂ precipitation occurs simultaneously [28]. The experimental conditions were set as follows: the reaction was conducted within a pH range of 8.0–10.0. The molar ratios were maintained at n(Mg):n(N) = 1–1.4 and n(P):n(N) = 0.9–1.3. The stirring speed was controlled within 120–240 rpm, the reaction time was set to 20–60 min, and the temperature was maintained at 15–35 °C.

The experimental conditions are shown in

Table A3. Experimental conditions for each test group (A1–F5), showing variations in pH, molar ratios (n(Mg):n(N), n(P):n(N)), stirring speed, reaction time, and temperature for MAP crystallization. Each experimental group corresponds to a specific parameter variation tested in the study.

No.	pH	n(Mg):n(N)	n(P):n(N)	Stirring Speed (rpm)	Time (min)	Temperature (°C)
A1	8	1	1	150	30	30
A2	8.5	1	1	150	30	30
A3	9	1	1	150	30	30
A4	9.5	1	1	150	30	30
A5	10	1	1	150	30	30
B1	9.5	1	1	150	30	30
B2	9.5	1.1	1	150	30	30
B3	9.5	1.2	1	150	30	30
B4	9.5	1.3	1	150	30	30
B5	9.5	1.4	1	150	30	30
C1	9.5	1	0.9	150	30	30
C2	9.5	1	1	150	30	30
C3	9.5	1	1.1	150	30	30
C4	9.5	1	1.2	150	30	30
C5	9.5	1	1.3	150	30	30
D1	9.5	1	1	120	30	30
D2	9.5	1	1	150	30	30
D3	9.5	1	1	180	30	30
D4	9.5	1	1	210	30	30
D5	9.5	1	1	240	30	30
E1	9.5	1	1	150	20	30
E2	9.5	1	1	150	30	30
E3	9.5	1	1	150	40	30
E4	9.5	1	1	150	50	30
E5	9.5	1	1	150	60	30
F1	9.5	1	1	150	30	15
F2	9.5	1	1	150	30	20
F3	9.5	1	1	150	30	25
F4	9.5	1	1	150	30	30
F5	9.5	1	1	150	30	35

, and the experimental procedure for MAP precipitation is shown in Figure A1.

2.3.2. Machine Learning Models

To evaluate the predictive performance of different machine learning approaches for the nitrogen removal efficiency (Φ, %) (raw data are listed in

Table A4. Triplicate nitrogen (N) removal efficiency (%) measurements for each experimental condition in Table A3.

No.	N Removal (%)-1	N Removal (%)-2	N Removal (%)-3
A1	87.7332	86.7154	87.5769
A2	90.0473	88.9712	90.1234
A3	91.0124	91.1908	89.7693
A4	95.1281	93.9163	94.3044
A5	95.4836	95.6868	96.5900
B1	91.5708	92.2728	92.9749
B2	94.2192	94.8262	95.4331
B3	96.0131	96.9656	97.9180
B4	96.6214	97.2976	97.9737
B5	95.5013	96.3917	96.2822
C1	91.2695	91.9950	92.7205
C2	93.0769	93.9505	94.8242
C3	93.8342	94.3003	94.7663
C4	95.4077	95.5266	94.2454
C5	94.0481	94.7256	95.4032
D1	93.9089	95.2225	96.5361
D2	93.9192	94.9542	95.9892
D3	92.3916	93.4934	93.5953
D4	92.2981	92.3742	91.4503
D5	90.4435	91.0191	91.6715
E1	92.1112	92.6441	93.1770
E2	92.0842	92.8235	93.5627
E3	93.1244	93.7442	94.3640
E4	94.3803	93.0267	94.0731
E5	94.7114	93.0746	94.4377
F1	92.4551	92.1105	91.7658
F2	93.2710	93.7442	93.2174
F3	94.2298	94.5693	94.9087
F4	93.9234	94.3421	94.7608
F5	95.4491	95.8965	95.3421

), six commonly used regression models were employed. Each model offered unique strengths in capturing linear or non-linear patterns in the data. These models were constructed using the Python programming language and the scikit-learn library.

Linear regression (LR) is a foundational statistical technique that models the linear relationship between independent variables and a continuous dependent variable. Despite its simplicity and interpretability, LR often faces limitations in handling non-linearity and multicollinearity [29].

Polynomial regression (PR) enhances the flexibility of linear regression by incorporating higher-degree predictor terms, allowing for the modeling of curved trends. In this study, second-degree polynomial features were used to capture the interactions and quadratic effects in the input space [30].

Ridge Regression is a regularized version of linear regression that introduces an L2 penalty term to reduce the impact of multicollinearity among the predictors. It improves the stability of a model by shrinking the coefficients while retaining all the input variables [31].

Lasso Regression, or the Least Absolute Shrinkage and Selection Operator, applies an L1 penalty to perform both regularization and feature selection. By forcing some coefficients to exactly zero, it results in simpler and more interpretable models, particularly when the predictor set is large [32].

Random Forest is a robust ensemble learning method that builds multiple decision trees using bootstrapped data and aggregates their outputs [33]. It handles non-linear relationships and interactions effectively and is well-suited to complex environmental modeling datasets with heterogeneous features [34].

Gradient Boosting is another ensemble method that builds decision trees sequentially, with each new tree trained to correct the prediction error of the previous ensemble. It often achieves higher accuracy than other models but is more sensitive to overfitting and the hyperparameter selection [35].

2.4. Analysis Methods

2.4.1. Nitrogen Removal Efficiency (Φ, %)

The efficiency of nitrogen removal (Φ, N removal (%)) was calculated based on the difference between the initial and final nitrogen concentrations in the wastewater samples, normalized by the initial concentration. This metric is commonly employed as a benchmark for evaluating the efficiency of the MAP precipitation method in removing ammonia nitrogen.

Nessler’s reagent and sodium potassium tartrate (KNaC₄H₄O₆·4H₂O) were used as chromogenic and masking agents, respectively, to measure the ammonia nitrogen concentration. The residual N concentration was measured using an Ultraviolet–Visible (UV–Vis) spectrophotometer at 420 nm (TU-1901, Beijing General Instrument Co., Ltd., Beijing, China).

Using Equation (2), the N removal (%) (Φ, %) was calculated from the difference between the initial (initial m(NH_4⁺-N), mg/L) and remaining concentrations (residual m(NH_4⁺-N), mg/L). The initial NH₄-N concentration was 560 mg/L.

$$\Phi \left(\%\right)={\displaystyle \frac{\left(m{\left(\mathrm{N}{\mathrm{H}}_4^{+}-\mathrm{N}\right)}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}-m{\left(\mathrm{N}{\mathrm{H}}_4^{+}-\mathrm{N}\right)}_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}}\right)}{\left(m{\left(\mathrm{N}{\mathrm{H}}_4^{+}-\mathrm{N}\right)}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}\right)}}\times 100\%$$

(2)

2.4.2. Calculation of MAP Purity

If the release of ammonia nitrogen as ammonia gas is not considered, all the removed ammonia nitrogen is converted into MAP precipitate and fixed. The purity of the MAP (P_MAP, (%)) in the obtained precipitate was calculated as the ratio of the theoretical mass of MAP (corresponding to the molar quantity of nitrogen removed) to the total mass of the precipitate.

$$P_{MAP}\left(\%\right)={\displaystyle \frac{{\left[N\right]}_0\times \Phi }{1000\times 14}}\times {\displaystyle \frac{M_{struvite}}{m_p}}\times 100\%$$

(3)

where [N]₀ is the initial N concentration, Φ is the N removal efficiency, m_p is the total mass of the precipitate, and M_struvite is the molar mass of the struvite.

2.4.3. Morphological Analysis Methods

X-ray diffraction (XRD) patterns of the precipitates were evaluated using an X-ray diffractometer (D8 advance, Bruker Corp., Germany) with a Cu Kα tube (λ = 1.5406 Å, 40 kV, 30 mA) scanned over a 2θ range of 10–70° at a rate of 10°/min. The morphology and structure of the sample were studied using a Gemini SEM 300 scanning electron microscope (SEM-EDS, ZEISS Corp., Jena, Germany) at the experimental facility in Shenyang City, Liaoning Province, China.

2.4.4. Model Development

Data analysis and computations were conducted using Python (version 3.12.4) in a Jupyter Notebook (version 7.0.8) environment. The following libraries were employed: Pandas (version 2.2.2) for data loading and preprocessing, NumPy (version 1.26.4) for numerical operations, and Scikit-learn (version 1.4.2) for machine learning tasks, including data scaling (StandardScaler), feature transformation (PolynomialFeatures), data splitting (train_test_split), and model training with linear models (Linear Regression, Ridge, Lasso) and ensemble methods (RandomForestRegressor, GradientBoostingRegressor). Model performance was evaluated using Scikit-learn metrics (r2_score, mean_squared_error, mean_absolute_error). Visualizations were generated using Matplotlib (version 3.8.4).

1.

Dataset Size

The dataset consists of 90 samples, corresponding to the experimental conditions detailed in Table A4. This dataset was derived from controlled experiments simulating high-concentration ammonia nitrogen wastewater treatment, ensuring a representative range of pH, molar ratios (n(Mg):n(N), n(P):n(N)), stirring speed, time, and temperature for model training and validation.

2.

Handling of Outliers

No specific outlier removal was performed, as the experimental data in Table A4 were pre-validated to ensure consistency. Standardization using “StandardScaler” was applied to normalize the feature distributions, mitigating the impact of potential outliers.

3.

Data Splitting Strategy

The dataset was split into training and testing in a 80/20 ratio, consistent with prior studies [36]. And the dataset sets used “train_test_split” with a fixed random state of 42 to ensure reproducibility.

4.

Feature Selection Rationale

Features including pH, n(Mg):n(N), n(P):n(N), stirring speed, time, and temperature were selected based on their known influence on nitrogen removal efficiency, as informed by experimental design and prior literature [37]. No automated feature selection was applied.

5.

Model hyperparameters

Linear Regression and Polynomial Regression (degree = 2) used default settings; Ridge Regression with alpha = 1.0, Lasso Regression with alpha = 0.1, Random Forest with n_estimators = 100 (optimized via GridSearchCV with n_estimators = [50, 100, 200], max_depth = [None, 10, 20], min_samples_split = [2, 5]), and Gradient Boosting with n_estimators = 100. Default values for other parameters. Hyperparameter tuning for Random Forest was conducted using 5-fold cross-validation via GridSearchCV to minimize mean squared error [38].

6.

Cross-Validation Procedure

A 5-fold cross-validation procedure was implemented during hyperparameter tuning with GridSearchCV to assess model performance robustly across different data subsets. This choice of 5 folds balances computational efficiency and reliability, aligning with common practices in machine learning studies [39].

The following metrics were used to assess accuracy and generalization: $R^2$ (coefficient of determination), RMSE (root mean square error), and MAE (mean absolute error) were calculated to assess the model’s accuracy and generalization ability. After validation, the best-performing model was used to evaluate the feature importance, identifying which factors most significantly influenced HANW nitrogen removal.

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-{\hat{\mathrm{y}}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(4)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\left(y_i-{\hat{\mathrm{y}}}_i\right)}^2}$$

(5)

$$MAE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}\left|y_i-{\hat{\mathrm{y}}}_i\right|$$

(6)

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} R^2=1-{\displaystyle \frac{\left(1-R^2\right)\left(n-1\right)}{n-k-1}}$$

(7)

where $y_{\mathrm{i}}$ is the observed value, ${\hat{y}}_{\mathrm{i}}$ is the predicted value, $\stackrel{\text{-}}{y}$ is the mean of the observed values, $R^2$ is the coefficient of determination, $n$ is the number of samples, and $k$ is the number of predictors.

3. Results

3.1. Effects of Parameters on N Removal Efficiency and Precipitate Characteristics

3.1.1. Effects of pH

The pH is an important parameter in MAP crystal formation, affecting the existing forms of phosphate and ammonia nitrogen. To investigate the effect of pH, experiments A1–A5 were conducted, and the results are shown in Figure 1.

As shown in Figure 1a, the nitrogen removal efficiency increased with pH, reaching a peak of 95.92% at pH 10. The observed trend is primarily attributed to the enhanced deprotonation of ammonium and phosphate species at higher pH levels [40]. Specifically, as the pH increased, ammonium was converted to NH₃, and the conversion of phosphate to H₂PO_3⁻ and H₂PO_4⁻ was inhibited [41,42]. Simultaneously, the concentration of the residual nitrogen in the solution decreased with an increasing pH, further confirming more efficient N removal. The MAP purity remained above 94% across all pH levels, indicating selective formation of struvite over competing phases such as magnesium hydroxide or amorphous phosphate salts. The mass of the precipitate also increased with the pH, reflecting the greater extent of crystallization and possibly the formation of larger or more densely packed crystals.

The phase identity of the precipitates is confirmed by the XRD patterns in Figure 1b, which show strong agreement with the reference pattern of struvite (JCPDS 77-2303). The characteristic peaks at 2θ angles of approximately 15.79°, 20.85°, and 33.27° are present in all the samples, confirming MAP formation under all the tested pH values [43]. The increase in the peak intensity at pH 9.5 indicates improved crystallinity, likely due to optimal ion availability and reduced solubility under these conditions. However, an overly high pH (>10) may also favor the co-precipitation of magnesium hydroxide, which could interfere with MAP’s purity and morphology. Variations in the relative intensity of reflections within the crystalline material’s X-ray diffraction (XRD) pattern indicate alterations in its texture and, more precisely, the size of its crystals [44].

The SEM images in Figure 1c–g provide visual evidence of pH-dependent morphological changes. At lower pH (8.0–8.5), the precipitates exhibit irregular and poorly defined crystal shapes, suggesting limited nucleation and growth due to insufficient supersaturation. As the pH increases to 9.0 and 9.5, the crystals become more uniform and elongated, exhibiting the typical rod- or needle-like morphology of well-formed struvite, similar to the figures in previous studies [23,45]. These structures indicate favorable growth conditions and stable crystallization. At pH 10.0, although the crystals’ quantity increases, their morphology becomes more aggregated and disordered, possibly due to excessive nucleation rates or incorporation of secondary phases such as amorphous magnesium phosphate.

Combined with the results shown in Figure 1a,c–g, this demonstrates that although the N removal efficiency (%) is the highest at pH 10.0, the oversized particles and lower purity of the MAP crystals make them unsuitable for subsequent application [46]. From the perspective of the N removal efficiency and MAP characteristics, it is most appropriate to set the pH of the reaction system to 9.5.

3.1.2. Effects of n(Mg):n(N)

To explore the impact of the n(Mg):n(N) molar ratio, experiments B1–B5 were carried out, with the outcomes presented in Figure 2.

As shown in Figure 2a, the nitrogen removal efficiency increased from 92.27% to 97.3% as the n(Mg):n(N) molar ratio rose from 1.0 to 1.3, as an appropriate excess of Mg²⁺ promotes complete precipitation by driving the equilibrium reaction forward [47]. However, when the ratio exceeded 1.3, further addition of Mg²⁺ resulted in a decline in both nitrogen removal efficiency and MAP purity, likely due to the formation of secondary phases such as Mg(OH)₂ or amorphous magnesium phosphate. This is consistent with previous studies showing that excessive Mg²⁺ increases the risk of non-struvite precipitates, reducing the selectivity of the crystallization process [48].

The XRD patterns shown in Figure 2b confirm that struvite was the main crystalline phase under all the tested conditions, with strong diffraction peaks aligned with JCPDS 77-2303. The increase in the peak intensity up to n(Mg):n(N) = 1.3 reflects improved crystallinity, while the emergence of peak broadening at 1.4 suggests a decrease in the product quality and possible interference from amorphous or poor crystalline impurities.

The SEM images shown in Figure 2c–g reveal morphological evolution with varying n(Mg):n(N) ratios. At low ratios (1.0–1.1), the crystals’ development is limited, and they are needle-shaped. With increasing Mg²⁺ (1.2–1.3), better-defined rod-like crystals are observed, characteristic of struvite formed under favorable growth conditions. At 1.4, particle aggregation and morphological distortion are evident, likely caused by accelerated nucleation and the presence of undesired phases, as previously reported in high-Mg systems [49].

In summary, controlling the n(Mg):n(N) molar ratio is critical for balancing the nitrogen recovery efficiency and product quality. An optimal ratio of n(Mg):n(N) = 1.3 ensures high removal efficiency and well-crystallized MAP with minimal secondary-phase contamination.

3.1.3. Effects of n(P):n(N)

To investigate the effect of the n(P):n(N) molar ratio, experiments C1–C5 were carried out, with the outcomes presented in Figure 3.

As shown in Figure 3a, the nitrogen removal efficiency peaked at n(P):n(N) = 1.2 with a value of 95.06%. However, further increasing the phosphorus dosage did not improve the nitrogen removal but instead caused a decline in the MAP purity. This can be attributed to the presence of excess phosphate in the system, which may have favored the co-precipitation of amorphous or alternative phosphate phases [50].

The XRD patterns in Figure 3b confirm that all the precipitates exhibited the characteristic peaks of struvite, as identified using JCPDS card No. 77-2303. The peak heights at n(P):n(N) ratios of 0.9, 1.1, and 1.2 are significantly lower than those at n(P):n(N) ratios of 1.3 and 1.0, showing the better purity of MAP at n(P):n(N) ratios of 1.3 and 1.0 [43].

The morphological changes observed in the SEM images shown in Figure 3c–g are consistent with these trends. At the optimal ratio shown in Figure 3d, well-defined rod-like struvite crystals dominate, reflecting stable and controlled growth. Below this stoichiometric ratio (Figure 3c), the crystal formation is incomplete due to a limited phosphate availability. When phosphorus is in excess, as shown in Figure 3e–g, the precipitates become irregular and poorly developed, suggesting that supersaturation disrupts normal crystal development and promotes secondary nucleation.

MAP with an even size and shape distribution is more conducive to improving fertilizer performance [51]. It can therefore be concluded that n(P):n(N) = 1.0 is the optimum reaction condition.

3.1.4. Effects of Stirring Speed

In order to investigate the influence of stirring speed, a series of experiments (D1–D5) were performed, and the corresponding results are presented in Figure 4.

As shown in Figure 4a, the nitrogen removal efficiency reached its maximum of 94.95% at 150 rpm and declined monotonically thereafter with further increases in stirring speed. This trend suggests that moderate agitation enhances mass transfer and promotes homogenous supersaturation, thereby accelerating nucleation and improving precipitation efficiency. In contrast, excessive shear may interfere with the crystal growth process by reducing the particle residence time in the supersaturated zone and inducing partial redissolution of forming nuclei due to turbulence [52].

With increased stirring speed, the purity and quality of MAP showed limited variation. While higher speeds enhanced collision frequency and slightly increased the total precipitate mass, they may have compromised crystal uniformity.

XRD analysis, as shown in Figure 4b, confirmed that struvite (JCPDS No. 77-2303) was the dominant crystalline phase under all tested conditions. The most intense and well-defined diffraction peaks appear at 120 rpm, indicating that this stirring speed favors the formation of highly crystalline and phase-pure struvite. At higher speeds, broader and less intense peaks may reflect incomplete crystallization or the presence of amorphous content, potentially caused by unstable mixing conditions and localized supersaturation [53].

The SEM images shown in Figure 4c–g reveal distinct morphological changes associated with the stirring speed. At 120–150 rpm, as shown in Figure 4c,d, struvite displays well-developed rod-like or prismatic crystals with smooth surfaces and uniform sizes, indicative of a stable growth environment with a sufficient nutrient supply and minimal mechanical disruption. In contrast, images at higher stirring speeds (Figure 4f,g) reveal fragmented and irregular structures, likely resulting from shear-induced crystal damage. Based on these findings, 150 rpm is identified as the optimal stirring speed.

3.1.5. Effects of Reaction Time

To explore the impact of the reaction time, experiments E1–E5 were carried out, with the outcomes presented in Figure 5.

As shown in Figure 5a, the nitrogen removal efficiency increased with the reaction time, suggesting that the majority of NH_4⁺ was removed during this initial period due to rapid struvite precipitation.

Interestingly, the MAP purity followed a non-monotonic trend, increasing from 20 to 40 min and then decreasing slightly at longer reaction times. This pattern can be attributed to the competitive dynamics between crystal growth and secondary processes. During the first 40 min, supersaturation drives both nucleation and orderly growth of struvite, resulting in high purity. However, after 40 min, the accumulation of byproducts or possible incorporation of impurities may reduce the relative purity of MAP.

The precipitate mass showed a steady increase with time, consistent with ongoing crystal growth and continuous removal of ions from the solution. Even though the majority of NH_4⁺ was removed within 40 min, crystal growth continued as residual PO_4³⁻ and Mg²⁺ remained available, contributing to mass accumulation.

The XRD patterns shown in Figure 5b confirm that struvite (JCPDS 77-2303) was the dominant crystalline phase across all time points. The sharpest peaks are observed at around 50–60 min, indicating good crystallinity. As shown in the SEM images in Figure 5c–g, the overall crystal size decreases with an increasing reaction time. The morphology evolves from a standard X-shape to a branched structure and gradually develops into an irregular, needle-like form, When the reaction time is extended to 60 min, the crystals tend to regain a more defined and complete morphology, ultimately forming a standard X-shaped structure. These observations support the idea that the reaction time modulates the transition from nucleation- to growth-dominated crystallization [54]. The shape changes as the variation in supersaturation and reaction time [23,55]. Initially, rapid supersaturation promotes nucleation; with time, crystal growth results in larger, more ordered particles. However, excessively long reaction times may allow secondary processes to occur [56], which can compromise the MAP purity despite continued mass gains.

According to Figure 5a–g, a reaction time of 30 min not only maintains the relative integrity of the crystalline structure but also ensures denitrification.

3.1.6. Effects of Temperature

To explore the impact of the reaction temperature, experiments F1–F5 were carried out, with the outcomes presented in Figure 6.

As shown in Figure 6a, an increase in the reaction temperature from 15 °C to 35 °C had an impact on the N removal efficiency, MAP purity, and precipitate mass. Overall, with an increase in the reaction temperature, the N removal efficiency and precipitate quality gradually increased, and the MAP purity showed a trend of gradually decreasing.

The reaction temperature had no discernible influence on the nucleation kinetics but exerted a pronounced effect on ionic solubility and reaction rate constants [57,58]. In the range of 25–30 °C, a modest temperature rise enhanced the ionic diffusivity and accelerated both nucleation and crystal growth, resulting in increased nitrogen removal and stoichiometric completion of the reaction. However, temperatures above 30 °C may destabilize the process, potentially due to increased struvite solubility or formation of competing phases, causing a slight decrease in MAP purity. The precipitate mass increased steadily with temperature, which may be attributed to both enhanced reaction rates and co-precipitation of secondary compounds at higher temperatures.

The XRD patterns shown in Figure 6b confirm that struvite was the dominant crystalline phase across all the reaction temperatures. The highest peak intensity appears at 30 °C, corresponding to the optimal crystallization conditions. At 35 °C, a slight reduction in the peak sharpness may indicate lower crystallinity or the presence of a minor amorphous phase.

The SEM images shown in Figure 6c–g reveal temperature-induced changes in the crystal morphology. At 15 °C, the precipitate forms loose, irregular agglomerates with a jagged outline. When the temperature is increased by 5 °C to 20–25 °C, the particle size decreases significantly, with wedge-shaped crystals mainly being formed, similar with previous study [25]. Further heating to 25–30 °C shows the recovery of greater sizes, the formation of polyhedral rod-shaped crystals with smooth sides, and a narrow size distribution, indicating that the growth kinetics are optimal. At 35 °C, the crystals appear to be slightly roughened and begin to cluster, suggesting instability at elevated temperatures, possibly due to surface dissolution or re-nucleation [59].

In summary, 30 °C appears to be the optimal temperature for struvite recovery under the studied conditions, balancing high nitrogen removal, high MAP purity, and a desirable crystal morphology.

3.2. Regression Model Equations

In this study, the dependent variable is the product yield (y). The independent variables include: pH (x₁), the molar ratio of magnesium to nitrogen (x₂, n(Mg):n(N)), the molar ratio of phosphorus to nitrogen (x₃, n(P):n(N)), stirring speed (x₄, rpm), reaction time (x₅, min), and temperature (x₆, °C). The empirical regression equation for the Linear models is expressed as

Table 1. Empirical regression equations for linear models.

Model	Empirical Regression Equation
Linear Regression	y = 39.7778 + 4.4016x1 + 8.8262x2 + 4.9330x3 − 0.0312x4 + 0.0118x5 + 0.0939x6
Ridge Regression	y = 50.8501 + 4.1982x1 + 3.7786x2 + 1.2674x3 − 0.0341x4 + 0.0033x5 + 0.1108x6
Lasso Regression	y = 97.7371 + 0.0000x1 + 0.0000x2 + 0.0000x3 − 0.0284x4 + 0.0000x5 + 0.0103x6

Note: The coefficients for pH, n(Mg):n(N), n(P):n(N), and Time were regularized to zero by the Lasso model, indicating their minimal contribution to the predictive model.

.

Unlike the linear regression models, ensemble methods such as Random Forest Regressor and Gradient Boosting Regressor do not have a simple empirical regression equation. These models operate on a series of complex decision rules derived from the training data rather than a fixed set of coefficients. This non-parametric nature allows them to capture intricate, non-linear relationships in the data more effectively, as evidenced by their superior performance metrics.

3.3. Factor Importance Analysis for Nitrogen Removal

To better understand each operational factor’s relative importance to the nitrogen removal efficiency, a comparative analysis was conducted.

The model performance was evaluated using the $R^2$, RMSE, MAE, and adjusted $R^2$ (

Table 2. Comparison of regression model performance in predicting nitrogen removal efficiency on the test dataset. Metrics include coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and adjusted R². The Random Forest and Optimized Random Forest models showed the highest predictive performance based on R² and error reduction.

Model	${\mathit{R}}^2$	RMSE	MAE	$\mathbf{Adjusted} {\mathit{R}}^2$
Linear Regression	0.890061	0.838068	0.736886	0.830094
Polynomial Regression	0.890863	0.835005	0.730308	0.831334
Ridge Regression	0.889255	0.841135	0.735733	0.828849
Lasso Regression	0.865746	0.926121	0.767374	0.792516
Random Forest	0.910994	0.754074	0.651988	0.862445
Gradient Boosting	0.889809	0.839027	0.720085	0.829705
Optimized RF	0.908239	0.765655	0.650953	0.858187

). Among all the models, Random Forest achieved the highest $R^2$ (0.911) and lowest RMSE (0.754) and MAE (0.652), indicating the best overall predictive accuracy. The optimized Random Forest showed similar performance ($R^2$ = 0.908), while the linear, polynomial, Ridge, and Gradient Boosting models showed moderate performance ($R^2$≈ 0.89). Lasso Regression performed the worst, likely due to over-regularization. Given its superior results, the Random Forest model was selected to perform further analysis of the feature importance.

Table 3. Comparison of experimentally determined and model-predicted optimal conditions for nitrogen removal via magnesium ammonium phosphate (MAP) crystallization. Parameters include pH, molar ratios of magnesium and phosphorus to nitrogen (n(Mg):n(N) and n(P):n(N)), stirring speed (rpm), reaction time (min), and temperature (°C). Nitrogen removal efficiency (%) is provided as the performance outcome.

Parameter	Experimental Optimal Conditions	Model-Predicted Optimal Conditions
Experiment No.	27	—
pH	9.5	9.8
n(Mg):n(N)	1.3	1.2
n(P):n(N)	1.0	1.0
Stirring speed (rpm)	150	150
Reaction time (min)	30	30
Temperature (°C)	30	25
N removal (%)	97.97	96.98

compares the optimal nitrogen removal conditions obtained from historical experiments and model predictions. Under the experimental optimal conditions (No. B4) of pH 9.5, an n(Mg):n(N) of 1.3, n(P):n(N) of 1.0, stirring speed of 150 rpm, 30 min reaction time, and 30 °C, 97.97% nitrogen removal was achieved. The model predicted optimal conditions of a slightly higher pH (9.8) and lower temperature (25 °C), with the other parameters unchanged, yielding a predicted nitrogen removal efficiency of 96.98%. The similarity between the two sets of conditions confirms the model’s reliability in optimizing the operational parameters.

Figure 7a depicts the correlation between the actual N removal values and those predicted by the Random Forest model. The scatter points are closely aligned with the 1:1 reference line, indicating a strong concordance between the predictions and measurements. This visual alignment, corroborated by quantitative metrics, validates the model’s accuracy in simulating N removal behavior. Such reliability supports the model’s application in process optimization and predictive analysis.

Figure 7b illustrates the relative importance of the input features to the Random Forest model’s predictions. It can be seen that the pH was the most important factor affecting N removal, highlighting its key role in this process. The n(Mg):n(N) ratio, stirring speed, and temperature also affected N removal, listed in order of decreasing importance.

4. Discussion

4.1. Effects of Operational Parameters on Struvite Formation

The formation efficiency and quality of struvite precipitates are significantly influenced by multiple operational parameters. Among the tested conditions, pH, molar ratios, stirring speed, reaction time, and temperature collectively determined the extent of nitrogen removal, MAP purity, and precipitate yield.

Section 3.1 illustrates that pH plays a crucial role in shaping crystal morphology, resulting in reduced crystal size and a transition from thick, broad structures to sharper, more slender forms. When the pH value reaches 10.0, crystallinity decreases compared to that under neutral conditions, which is closely related to the lattice distortion caused by residual sodium ions in the system [60]. Additionally, excess Na⁺ in the alkaline medium can occupy lattice sites through competitive complexation, resulting in limited three-dimensional crystal growth [61].

The molar ratios of n(Mg):n(N) and n(P):n(N) both favored the development of a looser, more porous crystal structure. In particular, increasing n(Mg):n(N) promoted a transformation from X-shaped to rod-like crystals, while higher n(P):n(N) values induced a shift from wedge-shaped to rod- and plate-like morphologies. According to classical kinetic theory, when the molar ratio exceeds 1.2, nuclear acceleration exceeds the crystal growth rate, which is consistent with the “explosive nucleation” phenomenon described by the LaMer model [62].

Stirring intensity had a notable impact on the crystal integrity, causing fragmentation and size reductions. Higher stirring speeds resulted in smaller crystals with more apparent morphological disruption. Stirring impacts the thickness of the diffusion boundary layer, fluid shear intensity, and the distribution of ions at the crystal–solution interface [63]. Insufficient agitation leads to concentration polarization and localized supersaturation, favoring uncontrolled nucleation. Excessive agitation may create shear stress that disrupts critical nuclei before they mature into stable crystals [64].

A prolonged reaction time facilitates structural rearrangement and crystal maturation, leading to the gradual formation of well-defined X-shaped crystals. This process is often accompanied by an increased surface roughness, indicating continued crystal growth and morphological development.

Temperature also had a significant influence, with MAP crystals exhibiting diverse morphologies, such as serrated, X-shaped, plate-like, and cross-shaped structures. At higher temperatures, signs of morphological degradation were evident, suggesting that thermal damage to the crystals impacted their growth and stability.

4.2. Scalability Considerations for MAP Process in HANW Treatment

While the laboratory-scale experiments demonstrated effective nitrogen removal and high-purity struvite precipitation under optimized conditions, scaling the MAP process to pilot or full-scale applications for high ammonia nitrogen wastewater (HANW) treatment presents several challenges and opportunities. The optimized parameters may require adjustments to accommodate the complex matrix of HANW, which often contains high salinity, organic matter, or heavy metals [65]. These components could interfere with crystal nucleation and growth, potentially reducing struvite purity and yield [16]. Pre-treatment steps, such as filtration or chemical conditioning, may be necessary to mitigate these effects, increasing operational complexity and costs.

From an engineering perspective, reactor design and operational conditions must be tailored for larger scales. For instance, achieving uniform mixing in large reactors to avoid concentration polarization, as observed with insufficient agitation in lab studies, requires advanced impeller designs or continuous flow systems. Excessive shear stress from high stirring speeds, which disrupts crystal integrity in lab settings, could be amplified in industrial reactors, necessitating precise control systems. Additionally, thermal management becomes critical at scale, as temperature fluctuations could exacerbate morphological degradation, as noted in Section 3.1.

Economic feasibility is another critical factor. While lab-scale MAP processes demonstrate high nitrogen removal efficiency, the costs of reagents (e.g., Mg and P sources) and energy for mixing and heating must be evaluated for full-scale implementation. Pilot studies integrating cost–benefit analyses and real-world HANW compositions are essential to validate scalability. Future research should focus on optimizing reactor configurations and exploring low-cost reagent alternatives to enhance the practical applicability of the MAP process.

5. Conclusions

This study investigated the influence of key operational parameters on ammonia nitrogen removal via magnesium ammonium phosphate (MAP) crystallization. Nitrogen removal efficiency was most affected by pH, followed by the n(Mg):n(N) molar ratio and stirring speed. It increased with pH, peaked at a moderate n(Mg):n(N), and declined as the stirring speed increased. Morphologically, MAP crystals underwent notable changes under varying conditions. A higher pH, n(Mg):n(N), stirring speed, and reaction time resulted in reduced crystal size and a transformation from typical X-shaped to needle-like structures. Under an increasing n(P):n(N), the crystals’ size first decreased and later increased, with their morphology shifting from X-shaped to rod- and needle-like forms. The highest removal efficiency (97.97%) was achieved at pH 9.5, n(Mg):n(N) = 1.3, n(P):n(N) = 1.0, 150 rpm, 30 min, and 30 °C. Future studies should prioritize pilot-scale experiments to validate the scalability of the MAP process for HANW treatment, addressing engineering and economic challenges to ensure practical implementation.

This study provides a theoretical and technical basis for magnesium ammonium phosphate crystallization treatment of high-concentration ammonia nitrogen wastewater, demonstrating that it is an important means to achieve sustainable development. This approach supports resource circularity by transforming nitrogen into a valuable agricultural product, reducing reliance on synthetic fertilizers and mitigating environmental impacts like eutrophication. The optimized parameters enhance process efficiency, paving the way for scalable, low-impact wastewater treatment solutions, as discussed in Section 4.2. This work aligns with sustainable policies for pollution control and circular economy, offering socio-economic benefits through waste valorization.