Pathogen Forecast Model Version 1.0: A Process-based Five-Day Beach Advisory Forecast System for the San Diego/Tijuana Region

Abstract

The San Diego/Tijuana (US/Mexico) coastal region has immense water quality issues due to untreated wastewater sources. Regular ddPCR Enterococcus sampling provides same-day water quality information for beachgoers but does not provide forecasts for decision making. We describe the Pathogen Forecast Model (PFM) V1.0, providing daily 5 day forecasts of swimmer illness risk along 30 km of shoreline on a web dashboard. PFM is based upon a hydrodynamic model using forecasted waves, winds, offshore currents, and Tijuana River flow as inputs. PFM accurately forecasts tides and waves and forecasts depth-averaged shelf currents moderately well. Modeled untreated wastewater concentration tracer C evolves with a constant decay rate representing Norovirus. Over 13 km of shoreline, forecasted C and observed Enterococcus E are moderately log-correlated (≈0.5) for day-1 and -5 forecasts. The C and E power-law relationship has an exponent of ≈0.5, possibly suggesting the tracer decays 2× slower than Enterococcus DNA, among other potential explanations. Using the cutoff C = 10^–5, PFM classified advisory exceedances with nearly 80% raw accuracy, high sensitivity, and lower specificity. This is the first forecast model that resolves estuarine, surfzone, and shelf regions. PFM is a useful public health tool and is constantly being improved.

Introduction

Untreated wastewater entering the ocean has large impacts on the San Diego/Tijuana (US–Mexico) regional beaches. Regional beaches such as Imperial Beach Pier, Silver Strand, and Coronado (Avenida Lunar) were under beach advisory the great majority of the time between May 2022 and April 2025, with subsequent impacts on recreation, tourism, and public health. The coastal border region (Figure ), denoted the San Diego Bight, has two major sources of untreated wastewater directly entering the coastal ocean. The binational Tijuana River (TJR) is often laden with untreated wastewater due to infrastructure problems in Tijuana. , The TJR enters the TJR Estuary (TJRE) and subsequently the coastal ocean at the estuary mouth. TJRE outflow can be surfzone-trapped and advected alongshore by relatively strong alongshore currents. It can also extend ∼10 km offshore as a plume. − A second source of untreated wastewater is at Punta Bandera (PTB10 km south of the border), where between 20 and 40 million gallons per day (MGD) of untreated wastewater from the City of Tijuana’s San Antonio de los Buenos (SAB) plant are released onto the shoreline year-round. Waves are an important driver of surfzone alongshore currents and thus are also important for alongcoast transport of untreated wastewater. During south swell (waves incident from the south) events, SAB/PTB plumes can be transported northward 30+ km in narrow, shoreline-attached plumes. , Water from these sources is associated with elevated fecal indicator bacteria, pathogenic bacterial genera associated with wastewater, , viruses, and chemical contaminants. , Through surfzone wave breaking, some of these bacteria and contaminants can become aerosolized and be transported inland. ,,

1.

LV4 model domain and bathymetry (colorscale) for the San Diego Bight, including, from North to South, Pt. Loma, San Diego Bay, the Tijuana River Estuary (TJRE), and the US/Mexico border. Yellow circles with red outlines indicate freshwater and untreated wastewater sources of the Tijuana River (TJR) and the San Antonio de los Buenos outfall at Punta Bandera (SAB/PTB). Yellow circles with black outlines indicate freshwater sources of the Otay River (OR) and Sweetwater River (SWR). Gray circles with black outlines mark locations of the San Diego Bay tide gauge, a wave buoy (CDIP), and 3 current meter locations (SBOO, MM, and A2). Purple squares indicate key beach locations from south to north: Imperial Beach (IB), south end of Seacoast Drive, Silver Strand State Beach (SS), and Coronado, Avenida Lunar (CO).

Due to these impacts, the San Diego County Division of Environmental Health and Quality, Beach and Bay Water Quality Monitoring Program performs regular (near daily) regional water quality testing. Although culture-based fecal indicator bacteria (FIB) testing has been performed for decades, since May 2022, the focus has been on ddPCR-based Enterococcus measurements with a beach advisory posting for samples exceeding 1413 copies/100 mL. Enterococcus is a fecal indicator with multiple possible sources, and in this region, elevated levels are often associated with untreated wastewater impacts. The ddPCR method measures Enterococcus DNA gene copies per volume, whether viable or not, whereas culture-based methods measure viable organisms under assay conditions. Many pathogens responsible for swimmer illness, such as Norovirus, are more persistent in the environment but harder to detect than Enterococcus. Enterococcus dies rapidly in the sunlight with short 1–6 h inactivation time scales, − whereas its DNA decay rates are considerably slower. In this sunny region, it can take many days for untreated wastewater to advect through the San Diego Bight. Thus, in this region, ddPCR-based measurements of Enterococcus are a suitable indicator of a fecal source and swimmer risk. In addition, ddPCR methods have the advantage of a faster turnaround time than culture-based methods, resulting in more rapid (same-day) beach advisory postings. The result of the frequent ddPCR sampling is the same-day knowledge of regional beach water quality, enabling bathers and beach managers to make effective decisions for a given day. However, it does not provide knowledge for future (5 day) public decision making. For example, midweek, a family cannot plan for a weekend beach trip. What is needed is a skillful forecast model, akin to weather models, for both pathogen concentrations and beach advisories based upon pathogen risk models.

Statistical models using various techniques (least-squares to neural networks) have successfully predicted beach advisories in freshwater , and ocean environments. , Such models use past water-quality data together with environmental data (e.g., rainfall, streamflow, waves, and tides), are tuned to individual locations, and are largely used in nowcast or hindcast modes. Statistical models were used to make 1 day forecasts in freshwater systems , and three-day forecasts at two California beaches. Process-based models solve coupled hydrodynamic and pathogen or FIB tracer equations with varying degrees of complexity and have been applied to freshwater, estuary, and coastal ocean systems largely in hindcast mode. , These models also usually do not include the relevant processes driving estuary, surfzone, and shelf currents. For example, some models are depth-integrated, whereas others are depth-resolving but do not include the full wave forcing terms. Alternatively, some models that include wave forcing do not resolve exchange with the inner shelf.

To study the potential benefits of the proposed San Diego/Tijuana region wastewater infrastructure projects, Feddersen et al. implemented a coupled wave and 3D circulation model that resolves the TJRE, surfzone, and shelf spanning 15 km cross-shore and 36 km alongshore, capturing both untreated wastewater point sources, at horizontal grid resolutions sufficient to resolve surfzone and estuarine processes. A model tracer represented untreated wastewater concentration with a decay time scale representative of norovirus. By simulating the year 2017, they showed that the SAB/PTB plume could advect northward in a narrow (a few 100 m wide) band along the coast during south swell events and heavily impact regional beaches. This highlighted the importance of waves to the alongshore transport of the untreated wastewater. Infrastructure projects that addressed SAB/PTB and moderately addressed the TJR gave the best benefits in terms of reduced modeled human illness and reduced modeled beach advisories. Here, we present a forecast model, the Pathogen Forecast Model (PFM), for untreated wastewater concentration and beach advisories in the San Diego/Tijuana coastal ocean that is based upon the Feddersen et al. hindcast model.

Ocean wave and circulation models have been used in forecast mode at a range of length-scales. Wave forecasting on a global scale is performed by NOAA and ECMWF. Regional scale wave forecasting is also performed, for example, in the Southern California Bight. , HYCOM and NEMO run by ECMWF are examples of global ocean circulation models that provide 10 day forecasts at 1/12 degree horizontal resolution. The USGS maintains a forecast system based upon a coupled wave-circulation model over the US East and Gulf coasts at 5 km resolution. Regional-scale forecast models at O(1) km resolution have been implemented that do not include wave coupling. − However, no coupled wave and circulation forecast system exists that includes the dynamics and small-scale O(10) m features of the surfzone, estuary, as well as the continental shelf.

Here, we report on version 1.0 of the PFM that provides 5 day forecasts of swimmer illness risk along 30 km of the San Diego/Tijuana shoreline. In PFM V1.0, untreated wastewater is represented as a tracer with a fixed decay constant representative of Norovirus. We compare TJR flow rates to the NOAA National Water Model forecast, and then compare our forecast of tides, waves, and currents to observations. These hydrodynamic parameters provide strong controls on the transport and dilution of untreated wastewater. We subsequently compare the forecast of untreated wastewater concentration to ddPCR Enterococcus measurements for both the first and fifth forecast day across a region spanning 13 km alongshore. We next examine PFM’s skill in forecasting beach advisories for both day-1 and day-5 forecasts. Lastly, we discuss the PFM V1.0 skill, compare PFM to a statistical water quality forecast, and outline our plans for PFM V2.0.

Methods

Forecast Model Description: Levels 1–4

The PFM consists of four nested ocean circulation simulations, denoted level 1 (LV1) to level 4 (LV4). Three parent domains (LV1–LV3; Figure S1) are one-way nested from LV1 to LV3 using the Regional Ocean Modeling System (ROMS) that does not include waves and provide boundary conditions for a high-resolution child domain of the US/Mexico border region (LV4; Figure ). The LV4 model includes waves by coupling ROMS to the phase-averaged SWAN (Simulating WAves Nearshore) wave model, via the Coupled Ocean-Atmosphere-Wave-Sediment Transport (COAWST) modeling system. , The wave coupling is crucial to resolve alongshore transport of untreated wastewater within the surfzone. Much of this configuration follows Wu et al. See the Supporting Information for more information on LV1-LV3. At all model levels, atmospheric forcing and boundary conditions are applied hourly, and model output is saved at 1 h intervals.

Level 4 (LV4): COAWST

The highest-resolution model domain (LV4) resolves the San Diego Bight in detail (Figure ), the shelf, surfzone, and the TJRE. The domain extends from Pt. Loma to south of Punta Bandera (PTB), including San Diego Bay and the TJRE (Figure ). The LV4 grid contains 486 × 1142 horizontal cells (15 km roughly cross-shore × 36 km roughly alongshore). The cross-shore resolution varies from 110 m at the offshore boundary to 7 m near the TJRE mouth. Alongshore resolution similarly varies, with a finest resolution of 7 m near the TJRE mouth and 30 m resolution near SAB/PTB. The grid size is a trade-off between grid resolution and computational costs. On this domain, surfzone tracer plumes can advect nearly 30 km alongcoast, , indicating sufficient resolution to simulate these processes. Bathymetry is based on the NOAA 1/3 arc-second coastal digital elevation model. We use 10 terrain-following vertical levels, a reduction from 15 in Wu et al., to enable faster runs and enhanced stability at high TJR outflows. ROMS and SWAN are two-way coupled every 10 min, similar to other studies, , and more than sufficient to resolve the variability induced by the wave–current interaction. Wetting and drying are implemented on LV4, allowing for a temporally and spatially varying shoreline with a minimum depth of 0.35 m, a value slightly larger than the 0.25 m used in Wu et al. to help model stability during large TJRE outflow.

The LV4 SWAN configuration uses 25 frequency bins (0.04–0.29 Hz) and 42 directional bins spanning 145°–355° (nautical convention), capturing all realistic wave approach angles. High directional resolution of 5° is required to resolve wave transformation to the surfzone, particularly for steep-angled incident south swells that drive northward alongshore surfzone currents. LV4 wave spectra boundary conditions are applied hourly on the southern, western, and northern boundaries (Figure ). The spatially varying boundary conditions are given by the CDIP operational wave model, driven by the ECMWF global wave forecast (via an ECMWF research license) offshore of the Southern California Bight. The standard value for the wave breaking parameter γ = 0.5 is used.

LV4 includes four freshwater sources: the Sweetwater River (SWR), Otay River (OR), TJR, and the San Antonio de los Buenos outfall at Punta Bandera (SAB/PTB) (yellow circles, Figure ). Forecasted discharges for SWR and OR are taken from the National Water Model (NWM) using reach IDs 948070199 and 20331702, respectively. The forecasted TJR freshwater discharge into the TJRE is given by the NWM forecast Q _TJ using reach ID 20324441. Note, NWM is a hydrological model and does not capture wastewater infrastructure changes that affect TJR flow. For SAB/PTB, PFM assumes a constant freshwater discharge Q _PTB. Prior to 15-April-2025, Q _PTB = 2.2 m³s^–1, and thereafter Q _PTB = 2.0 m³s^–1, reflecting changes in the San Antonio de los Buenos treatment plant reported to us via EPA or IBWC officials. All freshwater sources are set to have zero salinity and temperature equal to the spatially averaged land surface air temperature over LV4 at the forecast time. At the sources, salinity and temperature are distributed uniformly in the vertical.

The model simulates untreated wastewater concentration C as a nondimensional passive tracer with values between 0 and 1, where a value of 1 indicates 100% untreated wastewater. The tracer is advected, horizontally mixed with a constant diffusivity of 0.5 m²s^–1, and vertically mixed according to the k-ϵ model vertical diffusivity. In addition, the tracer has a constant decay rate. Although untreated wastewater has a variety of pathogens, we focus upon Norovirus as it is abundant in untreated wastewater, is relatively long-lived in seawater, , and is highly infectious, driving swimmer illness risk estimates in quantitative microbial risk assessments. ,− We set the decay rate to published estimates of Norovirus persistence in marine waters. Following Feddersen et al., from 5-Dec-2024 to 10-July-2025, the tracer decay rate was set to 0.1 d^–1 (10 day decay time-scale), based upon meta-analyses. , After 11-July-2025, the tracer decay rate was set to 0.143 d^–1 (7 day decay time scale) based on human norovirus infectivity decay rates in dark water. ,

The tracer has sources at the TJRE and at Punta Bandera (red outlined yellow circles in Figure ) that are also distributed uniformly in the vertical. At Punta Bandera (SAB/PTB), the input untreated wastewater concentration C _PTB is fixed. Prior to 15-Apr-2025, based on Feddersen et al., C _PTB = 0.7 was used. Thereafter, C _PTB = 0.5 is used to reflect improvements in the San Antonio de los Buenos wastewater treatment plant. Where the TJR enters the TJRE, the input untreated wastewater concentration C _TJ is poorly understood. Here, we assume a relationship between C _TJ and river flow rate Q _TJ, following two methods. Both of these methods have errors that are largely unquantifiable. Prior to 15-Apr-2025, the method used in Feddersen et al., based upon a Tijuana wastewater budget, was applied, which assumes a constant C _TJ = 0.3 for Q _TJ < 1.8 m³s^–1, whereas for Q _TJ ≥ 1.8 m³s^–1, the untreated wastewater volume flux is constant at 0.54 m³s^–1 (dashed black line in Figure ). After 15-Apr-2025, C _TJ is determined from parametrization of the Biggs et al. observations (colored markers in Figure ). The parametrization (solid black line in Figure ) assumes constant C _TJ = 0.65 for Q _TJ < 2.25 m³s^–1, whereas for Q _TJ ≥ 2.25 m³ s^–1, the untreated wastewater concentration is

$$C_{\mathrm{T}\mathrm{J}}=0.045+0.605\times (2.25{\mathrm{m}}^3{\mathrm{s}}^{-1})Q_{\mathrm{T}\mathrm{J}}^{-1}$$ 1

The constants in (1) are ad-hoc tuned for consistency with the 2023 observations. , We do not know how appropriate (1), based on 2023 observations, is in 2025 (and beyond) with its infrastructure changes.

2.

Tijuana River (TJR) untreated wastewater concentration C _TJ versus TJR discharge Q _TJ based on observations from Biggs et al. (colored symbols). The dashed curve is the parametrization used in Feddersen et al. and used here until 15-Apr-2025. The solid curve is the Biggs et al. parametrization tuned to fit the Feb and May 2023 data and used here after 15-Apr-2025.

Forecast Information

PFM is an open-source Python-based codebase. Similar to MacCready et al., , PFM automatically downloads and preprocesses required boundary conditions, surface forcing, and, if cold-started, initial conditions. It then performs the LV1-LV4 simulations and postprocesses model output for the web dashboard. Five-day PFM forecasts are run daily at 01:00 local Pacific time and are generally completed by 06:00. The model forecast begins at 00:00UTC (either 17:00PDT or 16:00PST the prior day) and uses a restart file from the previous forecast as the initial condition. Occasionally, the forecast did not run, primarily a result of missing required model inputs from HYCOM or NWM. In such cases, the forecast is typically rerun later in the day, when the forecast product becomes available. PFM forecasts began on 5-Dec-2024, and we include output here until 31-Aug-2025 (270 total days). Overall, PFM did not provide a forecast 14% of the time, although forecast frequency increased in later spring 2025 as the system became more robust. See the Supporting Information for further information on the forecast. As forecast skill typically is reduced with time, we focus forecast skill analysis on the day-1 and day-5 forecasts.

Summary of Major Model Changes During the Forecast Period

We will be comparing PFM forecasts to observations. However, it is important to note that the forecast model has undergone many changes since the forecasts began on 5-Dec-2024, as described above. These changes will affect the statistical relationship between model and observed quantities. The results here evaluate an evolving operational system. Here, we enumerate the major changes to PFM since the forecast began up through 31-August-2025.

• 9-January-2025: switched the atmospheric forcing from NOAA CONUS to NOAA GFS and changed forecast length from 2.5 to 5 d.

• 12-March-2025: switched the atmospheric forcing from NOAA GFS to ECMWF.

• 15-April-2025: Changed the SAB/PTB discharge from 2.2 m³s^–1 to 2 m³s^–1 and SAB/PTB input untreated wastewater concentration from 0.7 to 0.5. This change reflected the improvement in the SAB wastewater treatment plant.

• 15-April-2025: Changed the algorithm that relates TJR untreated wastewater concentration C _TJ to input flow rate Q _TJ based on the Biggs et al. observations.

• 11-July-2025: The e-folding decay time-scale for untreated wastewater was changed from (1/10) d^–1 to (1/7) d^–1.

Observations

Enterococcus Observations

The County of San Diego (California, USA) performs extensive water quality testing of the regional beaches. The water quality metric we use is Enterococcus, a common gut bacterium, measured using digital droplet PCR (or ddPCR) in units of copies/100 mL. This method compares reasonably well with traditional culture-based Enterococcus methods such as Enterolert. , We represent the ddPCR Enterococcus concentration as E. We focus on three sites sampled by the County in the San Diego South Bay region: the southern end of the City of Imperial Beach (end of Seacoast Drive), the entrance of the Silver Strand State Beach, and Avenida Lunar in Coronado (purple squares in Figure ). These 3 locations are chosen for their large alongcoast separation (13 km) and because, at these locations, water samples are taken nearly daily and, on average, on 85% of days. Water samples are taken in the morning, usually within 20 min of 6:30 am, 7:30 am, and 8:00 am at the end of Seacoast Drive, Silver Strand State Beach Entrance, and Avenida Lunar in Coronado, respectively (see Figure ).

Hydrodynamic Observations

Sea level observations come from the NOAA San Diego Bay tide gauge at Broadway Pier (tide gauge, Figure ). Waves were observed from the Imperial Beach nearshore wave buoy operated by the Coastal Data Information Program (CDIP) located in 21 m water depth (CDIP in Figure ). The Datawell Directional Waverider buoy reports wave spectra and frequency-resolved directional moments as well as bulk (frequency-integrated) parameters such as significant wave height H _s, mean period T _m, and mean wave direction D _m at 30 min intervals. Finally, the IBWC operates a flow gauge on the TJR near the international border (TJR in Figure ), providing observations of Q _TJ.

We perform model-data comparison with three acoustic Doppler current profilers (ADCPs) denoted SBOO, A2, and MM (Figure ) to validate the forecast model currents. The SBOO ADCP is located at 32 m water depth and 5.8 km from shore near the South Bay Ocean Outfall. The MM (Mini-Mooring) ADCP is located in approximately 11 m water depth, 850 m offshore of the Imperial Beach Pier (MM in Figure ). The A2 ADCP is located at 12 m water depth, about 800 m from the shoreline and is situated between the TJRE and SAB/PTB. For all ADCPs, observed currents over the vertical were aggregated to hourly averages using centered windows to match the hourly model output. We then depth-averaged the currents by vertical averaging between the shallowest and deepest useable velocity bin. See the Supporting Information for more detail on the ADCP deployments.

Results and Discussion

PFM Web Dashboard and Shoreline Untreated Wastewater Forecasts

The PFM web dashboard (https://pfmweb.ucsd.edu), publicly released on 21-July-2025, provides animations and a time series of 5 day forecasts of untreated wastewater concentration C along the San Diego–Tijuana coast (Figure ). The upper left quadrant provides language options, information on the latest forecast, and instructions on how to control the animation. In the lower left, a map of the San Diego–Tijuana coastal ocean is shown with overlaid contours of forecasted percent C from a low of just 0.0005% to as high as 10% (Figure ). Note that C is given as a percentage. The user has the ability to play, pause, rewind, or step through the five-day forecast. The shoreline is colored red, yellow, or green, indicating estimated swimmer illness risk according to forecast shoreline C. Red represents high risk (C > 0.1% and illness risk of ≥10%), yellow for moderate risk (0.001%–0.1%), and green for low risk (C < 0.001% and illness risk of ≤1%). The illness probabilities are estimated as in Feddersen et al. Clicking on the four marked beach locationsPlayas de Tijuana, Imperial Beach Pier, Silver Strand State Beach Entrance, or Coronado (Avenida Lunar) reveals, in the upper right, a time series of forecasted C, together with the estimated swimmer risk. The lower right has explanatory text noting that the forecast is experimental, has errors, and is not an official advisory. Over the 9 forecast months, the typical transport pathways of the TJRE and SAB/PTB sources vary between the wet (December-April) and dry (May-August) seasons (Figure S2), consistent with Feddersen et al. (see the Supporting Information).

3.

PFM web dashboard located at https://pfmweb.ucsd.edu. The upper-left region has information on the latest forecast and the play button and slider. The lower-left region is a map with overlaid animation of forecasted untreated wastewater concentration contours (in percent). The upper-right region shows forecasts of shoreline untreated wastewater concentration at particular locations: Playas de Tijuana, Imperial Beach Pier, Silver Strand, and Coronado - Avenida Lunar. The lower-right region provides model information.

Comparison to Hydrodynamic Observations

River Flow Comparison

Here, we compare the NWM forecast TJR flow rate, which is used as our forecast input, to the measured (at the IBWC flow gauge, Figure ) TJR flow rate Q _TJ. Because untreated wastewater input to the TJRE is linked to the TJR freshwater input Q _TJ (Figure ), accurate forecasts of the flow rate Q _TJ are needed to make accurate forecasts of shoreline untreated wastewater concentrations. The observed hourly Q _TJ varies from 10^–2 m³s^–1 to 300 m³s^–1 (Figure ) with diurnal variability of $\approx 1{\mathrm{m}}^3{\mathrm{s}}^{-1}$ , associated with diurnal untreated wastewater generation in Tijuana. The daily averaged ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ (blue line in Figure ) reveals three distinct time periods. The first time period spans December 2024 to late January 2025, where ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ varies generally from 0.5–1 m³s^–1. The second time period spans late January to early Apr 2025, where ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ is consistently elevated and often with ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}>2{\mathrm{m}}^3{\mathrm{s}}^{-1}$ . This is due to untreated wastewater diversion into the TJR during repairs to the SAB wastewater treatment plant in Tijuana. During this time period, peaks of a few days’ duration up to ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}=80{\mathrm{m}}^3{\mathrm{s}}^{-1}$ occur due to rainfall events. The third time period spans mid-April through August 2025, where ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ are $<1{\mathrm{m}}^3{\mathrm{s}}^{-1}$ and largely decline toward the end of August with occasional weak rainfall events.

4.

Time series of the Tijuana River flow rate into the Tijuana River Estuary (TJRE): hourly IBWC observations Q _TJ (gray), daily averaged IBWC observations ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ (blue), and first-day of the NOAA National Water Model (NWM) forecast Q _TJ (red) used by PFM. Occasional gaps in Q _TJ reflect times when the NWM forecast was unavailable, or the PFM forecast could not run.

The first day (first 24 h) of the NOAA National Water Model (NWM) river flow forecast (Q _TJ ) has some general similarities with the observed daily averaged flow rate $({\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}})$ , but also significant differences (red in Figure ). Because Q _TJ is based on both recent and forecasted rainfall together with a watershed model, diurnal variations in Tijuana’s wastewater usage are not captured. During the first time period (December 2024 to late January 2025), the forecasted flow rate is nearly flat at $Q_{\mathrm{T}\mathrm{J}}^{(\mathrm{n}\mathrm{w}\mathrm{m})}=0.4{\mathrm{m}}^3{\mathrm{s}}^{-1}$ and is consistently less than ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ by about 0.2–0.5 m³s^–1. During the second time period (late January to mid-April 2025), Q _TJ has four rain-fall-induced peaks at largely similar times as ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ but with quite different magnitudes and durations. The overall elevated background ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ is not captured in Q _TJ , as the NWM does not account for Tijuana untreated wastewater infrastructure changes. During the third period (mid-April through August 2025), the Q _TJ forecast consistently overpredicts ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ by around 1 m³s^–1 in late April to 0.3 m³s^–1 in late August. This level of NWM forecast error is consistent with previous research.

Tides and Water Level Comparison

LV4 forecast water levels compare very well with observations at San Diego Bay NOAA tide gauge (Figure ), with an overall Wilmott skill score (WS) and correlation of 0.99 and a root-mean-square error (RMSE) of 0.1 m for nearly 3 m of variability for both the day-1 to day-5 forecast. Both diurnal and semidiurnal components compare well, with a slight underprediction in the semidiurnal band; harmonic analysis of sea surface height for the observations and forecast (day-1 or day-5) is performed for the dominant tidal constituents (O1, P1, K1, N2, M2, S2, K2) with amplitude ≥0.05 m using the T_TIDE package. Model diurnal tidal constituents (O1, P1, and K1) exhibit small-amplitude error (<5%) (Figure ). Semidiurnal tidal constituents (N2, M2, and K2) also have small-amplitude error (<10%), whereas the weakest semidiurnal component, S2, has 20% error. Overall, the model underestimates the amplitudes of the semidiurnal constituents. For each constituent, the phase difference Δθ is defined as the model minus the observed phase, such that positive Δθ implies that the model lags the observations. For tidal constituents except S2 and K2, |Δθ| is <3^°. For the relatively weak S2, the Δθ is −5° and −8.5°, for the day-1 and day-5 model forecasts, respectively, whereas for the weak K2, the Δθ is −7.5°. Overall, the small Δθ indicates the model well reproduces tidal phase within the LV4 domain. In the subtidal (time scales longer than 40 h) band for day-1 and day-5 forecasts, sea-surface variability is much smaller (0.3 m) and is underpredicted (r = 0.75, WS = 0.78, and RMSE = 0.08 m).

5.

Harmonic analysis for tidal constituents O1, P1, K1, N2, M2, S2, and K2 at the San Diego Bay tide station (Figure ): (a) observed (black) and day-1 forecast (red) amplitude, and (b) phase difference Δθ between modeled and observed. Only tidal constituents with amplitudes greater than 0.05 m are shown.

Wave Comparison

LV4 day-1 and day-5 forecasted waves compare very well with observations from the CDIP wave buoy (see Figure ). From December through March, waves are generally incident from the northwest and have significant wave height H _s varying from 1–4 m, mean wave period T _m between 5 and 15 s, and mean wave angle D _m mostly out of the west ≈270^° (Figure ). A transition occurs from May through August, where a wave field consists of swell events incident from the southwest together with smaller seas incident from the northwest, with H _s between 0.7–2 m and T _m generally <10 s. The CDIP wave buoy observed H _s, T _m, and D _m agree well with the first day forecast values (Figure ). Observed and day-1 forecasted H _s agree very well with a correlation of 0.93, a WS of 0.96, and an RMSE of 0.17 m. Observed and day-1 forecasted T _m also agree well, with a correlation of 0.81, a WS of 0.9, and an RMSE of 1.13 s. Similarly, the observed and day-1 forecasted mean wave direction D _m also agree well with a high correlation of 0.84, a WS of 0.88, and a low RMSE of 7 deg, just larger than the model 5 deg directional bin size. The skill metrics for day-5 forecasted waves are reduced only slightly relative to the day-1 forecasts (see Supporting Information).

6.

Time series of (a) significant wave height (H _s), (b) mean wave period (T _m), and (c) mean wave direction (D _m) at the Imperial Beach Nearshore CDIP wave buoy (Figure , CDIP). Observed and modeled values are shown in the solid black and red lines, respectively. Mean wave direction is the direction waves are coming from.

Depth-Averaged Currents Comparison

Hourly day-1 and day-5 forecast depth-averaged currents are extracted from the mooring locations analogous to the observations. In the Southern California Bight, low frequency (<1/40 cph) alongshelf currents are typically driven by alongshelf winds and pressure gradients, in contrast to the dominant wave-forcing in the surfzone. A principal axis analysis is performed to rotate the velocities into major (U _maj, oriented alongshore) and minor (U _min, oriented cross-shore) axes. The observed and modeled principal axes are nearly identical: 2° difference for SBOO and 4° difference for A2, with the exception of MM, which has a 29° difference; the latter is likely due to local bathymetric influences. Observed and day-1 forecast U _maj across the moorings are relatively consistent, with reduced forecast U _maj magnitude and correlations between 0.32–0.44, WS ≈0.6, and RMSE of 0.1 ms^–1 (Figure a,c,e). The observed U _maj skill metrics are higher at the shallower A2 and MM locations where the currents are stronger relative to the SBOO, whose currents are weaker and skill metrics are lower (Figure a,c,e). At SBOO and A2, persistent southward flow in late April and early August is poorly forecast. This is likely due to errors in the forecast alongshelf pressure gradients, which can be a significant driver of alongshelf flow in the Southern California Bight. , Observed and forecast U _min (Figure b,d,f) are considerably weaker than U _maj and are not related at any mooring. Overall, the forecast has greater skill in the semidiurnal and diurnal tidal bands than in the lower frequency subtidal (>40 h) band (Figure S3 and Supporting Information). The skill metrics for day-5 forecast U _maj are reduced only slightly relative to the day-1 forecasts (see Supporting Information).

7.

Time series of day-1 forecast (red) and observed (black) depth-averaged (left panels) major axis currents U _maj and (right panels) minor axis currents U _min. The rows represent different locations (see Figure ) (a,b) SBOO, (c,d) South of US-MX border (A2), and (e,f) Imperial Beach Mini Mooring (MM). In the left panels, water depth h and U _maj statistics of correlation r, Wilmott Skill Score (WS), and root-mean-square error (RMSE) are given.

Comparison of Forecasted Untreated Wastewater Fraction to ddPCR Enterococcus

We now examine the relationship between the shoreline PFM untreated wastewater concentration C and ddPCR Enterococcus concentration E. We assume that for pure untreated wastewater (C = 1), there is a fixed E, and thus, that E and C have a consistent relationship. Pathogen concentration in untreated wastewater varies seasonally, , and thus source E likely also varies, potentially adding noise to any E and C relationship. We evaluate PFM at three shoreline locations where regular ddPCR Enterococcus sampling occurs. From south to north, spanning 13 km, these locations are Imperial Beach: South End of Seacoast Drive (IB), Silver Strand State Beach Entrance (SS), and Coronado: Avenida Lunar (CO) (pink square, Figure ). Note that we do not have ddPCR Enterococcus measurements in Mexico, and so we cannot examine the forecast skill there. For each day of the PFM forecast, we average the model shoreline C over ±3 h centered at the sample collection time. Results do not depend upon ±1 h to ±6 h averaging duration. As the forecast usually starts at 00:00UTC, a 07:00 local time (in PST or PDT) sample collection is either at 14 or 15 h into the simulation for the first forecast day and at (14|15) + 24­(d – 1) h into the simulation for each subsequent forecast day, where d represents the forecast day. We denote the concentration on the nth forecast day as C _n . We evaluate PFM from 9 January (when 5 day forecasts began) to 31 August 2025. Recall that, during this time period, significant changes to the forecast model were made.

During this time, the ddPCR Enterococcus measurements span a large range from 90 copies/100 mL to 2 × 10⁶ copies/100 mL (Figure a,b). Generally, values are larger at the southernmost (IB) location and smallest farther north in CO. However, SS can be as large as 7.6 × 10⁵ copies/100 mL and CO as large as 4.3 × 10⁵ copies/100 mL, and at IB can be as small as 414 copies/100 mL. Similarly, over this time period and over both day-1 and -5 forecasts, the modeled untreated wastewater concentration C varies also over a very large range from 10^–7 to 10^–2 (Figure a,b), with generally weaker levels to the north farther from the sources. The C and E have significant and similar log correlations (defined as correlation of log₁₀(C) and log₁₀(E)) of r _log = 0.51 for the day-1 and r _log = 0.48 for the day-5 forecasts. When broken down by region, day-1 forecast r _log = {0.44, 0.51, 0.31} at IB, SS, and Coronado, respectively. For the day-5 forecast, the regional correlations are similar to r _log = {0.45, 0.47, 0.31}. At lower E levels (e.g., $<1413\mathrm{c}\mathrm{o}\mathrm{p}\mathrm{i}\mathrm{e}\mathrm{s}/100\mathrm{m}\mathrm{l}$ ), more C scatter is evident than at higher levels of E. Across all three regions, the best fit power-law relationship between C and E results in

$$C_1=10^{-6.46(\pm 0.17)}\times E^{0.57(\pm 0.042)}$$ 2

$$C_5=10^{-6.33(\pm 0.17)}\times E^{0.52(\pm 0.042)}$$ 3

for day-1 (C ₁) and day-5 (C ₅), where the values in parentheses represent the estimated standard deviation. The E exponents of 0.57 and 0.52 are similar and statistically indistinguishable. Similarly, the preceding constants (10^–6.46 and 10^–6.33) are similar and, based on the estimated standard deviation, also statistically indistinguishable. The moderate and consistent correlation and the consistent power-law relationship across forecast days indicate that PFM has reasonable skill across all five forecast days despite the significant model changes and TJR input errors (Figure ).

8.

Forecasted untreated wastewater concentration C versus ddPCR Enterococcus measurements E from 9 January 2025 to 31 August 2025: (a) day-1 forecast and (b) day-5 forecast, which starts 14 January 2025. The number of data points are (a) 528 and (b) 512. The logarithmic correlations are (a) r _log = 0.51 and (b) r _log = 0.48. The locations of Imperial Beach: End Seacoast Drive (IB), Silver Strand State Beach Entrance (SS), and Coronado: Avenida Lunar (CO) are represented by diamonds, circles, and triangles, respectively. The magenta dashed line represents the best-fit power law. The vertical dashed line represents the ddPCR Enterococcus beach advisory threshold of 1413 copies/100 mL. The horizontal lines at C = 10^–5 and C = 10^–3 separate regions identified in color as low (green), moderate (yellow), and high risk (red) for swimmer illness and correspond to a 1% and 10% estimated norovirus swimmer illness probability, respectively.

Forecasting Beach Advisories

Having a good statistical relationship between PFM untreated wastewater fraction and the San Diego County ddPCR Enterococcus measurements gives us confidence in PFM. However, this alone does not provide beachgoers with actionable information on whether to swim. San Diego County uses the threshold of E _a = 1413 copies/100 mL to make beach advisory decisions, as shown by the dashed vertical line in Figure . In the web dashboard, the forecast model has two thresholds: C = 10^–5 for moderate risk and C = 10^–3 for high risk. Here, we determine the optimal threshold value C _a for accurate forecasts of regional beach advisories. A predicted positive beach advisory occurs when forecasted C ≥ C _a, and a negative occurs when C < C _a. A true positive (TP) is a predicted positive together with E ≥ E _a, while a false positive (FP) is a predicted positive with E < E _a. Likewise, a false negative (FN) is a predicted negative with E ≥ E _a, and a true negative (TN) is a predicted negative with E < E _a. For a particular C _a, the fraction of accurate prediction γ is defined as the number of TP and TN divided by the total number of data points. We estimate γ as a function of C _a for both the day-1 and day-5 forecasts (Figure ).

9.

PFM-forecasted beach advisory accuracy fraction γ versus untreated wastewater concentration threshold C _a for (a) day-1 (blue circles) and day-5 (red diamonds) forecasts for the time period 9-January to 31-August 2025 across all 3 locations. (b) PFM-forecasted γ versus threshold C _a for Imperial Beach: End of Seacoast Drive (IB, blue), Silver Strand State Beach Entrance (SS, green), and Coronado: Avenida Lunar (CO, red). Day-1 and day-5 forecasts are indicated with circles and diamonds, respectively. The vertical dashed line is C _a = 10^–5.

For both forecast days, γ ≈ 0.8 for C _a ≤ 10^–5, indicating that for C _a values in this range, the model will forecast whether or not a beach would be under advisory correctly approximately 80% of the time (blue circles, Figure a) across the 3 regions. For C _a ≥ 10^–5, γ decreases with increasing C _a. We choose the threshold value C _a = 10^–5 as it is the C _a where γ begins significantly decreasing (from its maximum value of ≈0.8). The γ and C _a relationship is similar for the day-5 and day-1 forecasts (Figure a). This is consistent with the similar overall C and E relationship (2, 3) across day-1 and day-5 forecasts.

Overall, for all C _a, the day-1 and day-5 forecast γ are similar when split by location (Figure b), with day-5 γ generally slightly smaller than day-1 γ (compare same colored diamonds to circles in Figure b), similar to the overall γ (Figure a) and the C and E relationship (Figure ). At each location, the relationship between γ and C _a is similar; however, there are some differences. IB has γ = 0.9 for C _a ≤ 10^–5, and γ decreases rapidly for C _a larger than 10^–5. At SS, γ is variable between 0.85 and 0.80 for C _a ≤ 10^–5 and decreases rapidly for larger C _a. At CO, γ increases from 0.63 for C _a = 10^–7 to a maximum γ = 0.68 for C _a = 10^–5, before decreasing rapidly for C _a > 10^–5 (red in Figure b). At the three individual locations, C _a = 10^–5 is a good threshold value overall. With C _a = 10^–5, γ decreases quasi-linearly with alongcoast distance across these three locations. The pattern of elevated γ at IB that decreases farther north also indicates that forecast accuracy decreases with distance from the untreated wastewater sources, consistent with the decreased C and E correlation with distance from the sources.

Choosing the threshold of C _a = 10^–5 for all three locations, we now examine the distribution of TP, FN, FP, and TN at the three locations (Figure ). For the day-1 forecast (Figure a), at IB, TP dominates occurrence fraction at 0.89, FN and FP are 0.07 and 0.03, respectively, and TN is 0.01. At SS, the TP occurrence fraction is also high at 0.68, but reduced relative to IB; FN is 0.12, FP is 0.09, and TN is 0.10. At CO, the TP occurrence rate is 0.51, again lower than SS, and FN is 0.10, FP is 0.21, and TN is 0.17. FN and TN are much larger than the other locations. Thus, with distance from the two pollution sources, TP decreases, FN is relatively constant, and FP and TN increase. Generally, FP is either greater than TN or similar to TN. The day-5 forecast occurrence fraction (Figure b) is similar to that for day-1 (Figure a), again consistent with previous results.

10.

Occurrence fraction of beach advisory true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN) for the three locations of Imperial Beach: End of Seacoast Drive (IB), Silver Strand State Beach Entrance (SS), and Coronado: Avenida Lunar (CO) for (a) day-1 and (b) day-5 forecasts.

We next examined the sensitivity (Se) and specificity (Sp) of the PFM forecasts relative to ddPCR Enterococcus as well as positive (PPV) and negative (NPV) predictive values. Sensitivity, Se (specificity Sp), gives the probability of forecasting a positive (negative) beach advisory conditioned upon all positive (negative) beach advisories. PPV (NPV) gives the probability of true positive (negative) beach advisories conditioned over all positive (negative) forecasts. They are defined as

$$\mathrm{S}\mathrm{e}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$ 4

$$\mathrm{S}\mathrm{p}=\frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}$$ 5

$$\mathrm{P}\mathrm{P}\mathrm{V}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$ 6

$$\mathrm{N}\mathrm{P}\mathrm{V}=\frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}$$ 7

The sensitivity of PFM forecasts is 0.93, 0.85, and 0.83 at IB, SS, and CO, respectively, indicating that PFM forecasted most observed beach advisories. Similarly, the day-1 forecast PPV is high, with values of 0.97, 0.88, and 0.71 at IB, SS, and CO, respectively. Thus, if the model forecasts a positive, the likelihood of an actual positive E is high at all three locations, with the likelihood decreasing northward. Such information is actionable for a beachgoer who wants to bathe. Specificities of PFM day-1 forecasts are 0.33, 0.53, and 0.44 at IB, SS, and CO, respectively. The day-1 forecast NPV is 0.17, 0.47, and 0.62 at IB, SS, and CO, respectively. This implies that PFM is poor at forecasting when the beach should not be under advisory. This is consistent with increased C scatter for E < E _a (Figure ). This information is less actionable for beachgoers. At IB, the total positive rate is so high that always forecasting a beach advisory provides very high sensitivity but results in low specificity. The day-5 forecast sensitivity is slightly reduced at 0.89, 0.84, and 0.80 at IB, SS, and CO, respectively. The day-5 forecast specificity is higher at IB at 0.50 and similar to the day-1 forecast at SS and CO (Figure b). The day-5 forecast PPV and NPV are similar to the day-1 forecast values.

Discussion of Forecast Skill and Potential Model Improvements

Even with the significant changes in the model, error in TJR flow rate Q _TJ (Figure ), and moderate skill in forecasted shelf currents (Figures and S3), PFM has moderate skill between untreated wastewater fraction C and ddPCR Enterococcus observations E across a wet and dry season and 13 km of alongshore distance (Figure ). The skill is likely due, in part, to strong skill in forecasting waves and thus wave-driven flows within the surfzone. Even for perfect PFM forecasted currents, the error in Q _TJ (Figure ) must result in a reduced correlation between C and E and increased RMSE. The National Water Model (NWM) forecast Q _TJ has significant errors during both wet and dry seasons (Figure ), similar to NWM errors found in other locations. Thus, improving Q _TJ forecasts likely would improve forecast model skill. Often, the daily averaged flow rate ${\mathrm{Q̅}}_{\mathrm{T}\mathrm{J}}$ does not vary substantially from day to day, particularly during summer (dry) season. A statistical mix of a persistence forecast together with NWM when rain is forecasted should improve the RMSE of the forecast Q _TJ. This is an analogous approach to Chin et al., who developed a Long Short-Term Memory network to adjust the NWM forecast flow rate. We have implemented such an approach beginning on 1-September-2025, which will be reported elsewhere.

If C were to predict ddPCR Enterococcus E, then the best-fit power law exponent should be one. However, the exponent (≈0.55) is less than one, suggesting a relationship $C\propto \sqrt{E}$ . This implies that E decays twice as rapidly as C, for which there may be many potential causes. Here, we discuss two potential causes. The first is that the ocean’s hydrodynamic dilution is more rapid than the model’s. The second is that, via biotic and abiotic processes, the time-scale for Enterococcus DNA in the ocean is shorter than the model’s 10 d (7 d after mid-July) decay time-scale.

An additional potential reason for the imperfect C skill is that the forecasted shelf currents are weaker than observed and, particularly for the subtidal flow, not perfectly phased (Figures and S3). Errors in forecasted currents will result in errors in forecasted tracers as tracers are advected and dispersed by currents in the surf zone and shelf. The reason that observed currents are underpredicted is not known, but there are two broad possibilities. First, the forecasted winds (not shown here) are somewhat weaker than observed, which will result in weaker wind-driven currents. Another possibility is that the model bottom friction is too strong and that, by changing model bottom stress parameters, forecasted and observed currents could have similar magnitudes. Investigating this possibility will be reported elsewhere.

Contextualizing Process-based and Statistical Forecasts

We contextualize the process-based PFM 5 day forecast with the statistical-based 3 day forecast of Searcy and Boehm. In contrast to PFM’s use of forecasted inputs, Searcy and Boehm used environmental variables available only up to the forecast start time. They trained a suite of machine learning models at 2 California Beaches to forecast summer FIB exceedance probability (not FIB concentration). At both locations, the training and testing data rarely exceeded the regulatory standard (<13% and <8%). This is in contrast to IB, SS, and CO, where ddPCR Enterococcus exceeded the advisory threshold 96%, 80%, and 61% of the time, respectively. Across all models and forecast days, the Searcy and Boehm sensitivity rate was low, and the specificity was high. This is generally opposite of the sensitivity and specificity pattern of the PFM 5 day forecast spanning 13 km of coastline. The better sensitivity of PFM but worse specificity is in part due to the San Diego Bight beaches being far more contaminated than those forecasted by Searcy and Boehm.

Statistical and process-based forecast models have their own pros and cons. Statistical models are computationally efficient, particularly for prediction. Process-based models are much more computationally demanding, here requiring a dedicated computing cluster. Statistical models must be trained separately for each location, with large training data sets. Process-based models seamlessly make predictions everywhere in the model domain, relying on the model physics and biology. However, this requires that the physics and biology, as well as the forcing (flow rate, waves, etc.), be accurate. If the statistical relationship between environmental parameters and water quality changes, such as due to an infrastructure change, the statistical model must be retrained, requiring sufficient new data to be available. Process-based models do not have this issue as long as they are quantitatively informed of the changes.

Summary and Future Plans for PFM V2.0

The PFM version 1.0 makes 5 day forecasts of modeled-derived untreated wastewater tracer for the San Diego Bight coastal waters using a process-based model. Untreated wastewater concentration C is represented with a tracer with a constant decay rate. The forecast runs every day and is generally available on the web dashboard at 06:00 Pacific Time. We analyze the forecast output from December 2024 through August 2025, during which a number of significant model changes occurred. Thus, we are evaluating an evolving operational system. PFM forecasts tides and waves with high skill and forecasts depth-averaged shelf currents with moderate skill. The high wave forecast skill suggests the surfzone alongshore currents should be accurately forecasted. A key model input is the NOAA NWM forecast of the TJR flow rate, which has significant error relative to the observations. Over the analysis period and spanning 13 km of shoreline, the model C and observed ddPCR Enterococcus E have a power law relationship $C\propto \sqrt{E}$ and a moderate log correlation of ≈0.5. This suggests PFM is doing well given the model changes, the flow rate error, error in the forecast currents, as well as other sources of error. Using a threshold of C _a = 10^–5, the day-1 and day-5 forecasts predict reasonably well (78%) whether the beach advisory threshold for E will or will not be exceeded. Although the forecast sensitivity is quite high (>0.8), the specificity is only moderate at 0.3 to 0.5. PFM V1.0 is a useful public health tool for the communities of this highly impacted region. All PFM components are open-source, and the forecast components are based on global forecasts. Thus, it would be possible to adapt PFM to other locations impacted by untreated wastewater.

We close with our plans for PFM v2.0. Most weather and some ocean forecasting systems , use data assimilation so that their forecast initial conditions are consistent with available data. The Southern California Bight region has a plethora of oceanographic observationssuch as HF radar currents , which could be used to improve the forecast initial condition and thus the forecast itself. This is anticipated to improve shelf currents. The astute reader will have noticed that with the untreated wastewater tracer, we are not simulating a pathogen. PFM V2.0 will address this limitation with a Norovirus tracer. Pathogen persistence and infectivity depend upon environmental variables such as sunlight, water temperature, and salinity. ,,, The Norovirus model decay rate will depend upon environmental variables available in the COAWST model, with specific dependence based upon recent and ongoing work. Norovirus input concentration will be based upon the WastewaterSCAN wastewater surveillance program that reports near-real-time Norovirus (and other pathogen) concentrations entering many wastewater treatment plants around the country. ,, With PFM V2.0, we hope to make PFM an even more useful public health tool.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.6c04225.

• Additional model information and information on current meter deployments, on the forecasts of shoreline untreated wastewater, and comparison between model forecast and current meter data (PDF)

The authors declare no competing financial interest.