To reconcile the need to sustain river communities, water quality, and ecosystem services against the needs for water supply, daily usage, and power generation, to name a few, extensive understanding and accurate sediment estimates are required. However, sediment's nonlinearity and complex hysteresis characteristics make sediment prediction challenging. Furthermore, most of the past studies have focused solely on a local scale and simply ignored human-induced disturbances, land cover status, geological properties, etc. There lacked a systematic way to predict sediment contributions at large scales and high accuracy. This study developed models based on long short-term memory (LSTM) deep networks to predict SSC over 377 sites across the Contiguous United States (CONUS) with hydrometeorological forcing datasets from the DAYMET, Daily Surface Weather, and Climatological Summaries (Jan 1980 -- Dec 2019), streamflow from USGS, United States Geological Survey, sediment-related static attributes from GAGES-II, Geospatial Attributes of Gages for Evaluating Streamflow or non-sediment-related static attributes (simulated random vectors) as inputs. The model has been trained either basin by basin (called the local model) or for the entire CONUS (call the Whole-CONUS model) with the optimal period to cope with data availability and variations of each basin. The local models for 377 sites across the CONUS show much superior prediction performance compared to the Whole-CONUS model, in terms of a median of the Nash-Sutcliffe Error (NSE), Pearson's Correlation Coefficient (R), Coefficient of Determination (R2), Root Mean Square Error (RMSE, [mg/L]), and Bias [mg/L] of all sites equal to 0.687, 0.868, 0.727, 67.295, 1.379 respectively in testing. The result from the Whole-CONUS from all sites training (377 sites) with sediment-related attributes shows relatively lower statistical metrics (0.596, 0.806, 0.645, 79.303, and -3.403. for Nash-Sutcliffe Error (NSE), Pearson's Correlation Coefficient (R), Coefficient of Determination (R2), Root Mean Squared Error, and Bias respectively) and with random vector attributes ((0.508, 0.788, 0.621, 92.186, and -4.507. for Nash-Sutcliffe Error (NSE), Pearson's Correlation Coefficient (R), Coefficient of Determination (R2), Root Mean Squared Error, and Bias respectively). For ungauged site experiment, the Whole-CONUS with 284 training sites perform a satisfying performance in Suspended Sediment Concentration prediction with a 0.591, 0.810, 0.651, 84.598, and -1.669 for Nash-Sutcliffe Error (NSE), Pearson's Correlation Coefficient (R), Coefficient of Determination (R2), Root Mean Squared Error, and Bias respectively. With a Whole-CONUS that has been trained with 284 basins, it can predict the 93-ungauged sites with satisfactory performance (0.521, 0.865, 0.749, 79.129, and -0.211 for Nash-Sutcliffe Error (NSE), Pearson's Correlation Coefficient (R), Coefficient of Determination (R2), Root Mean Squared Error, and Bias) The results from this study suggest that on gauged sites, the locally-trained LSTM model is the best approach to estimate the SSC time series. In the ungauged sites, the Whole-CONUS can be considered an acceptable approach for expanding the availability of SSC data. The Whole-CONUS model has been trained either with sediment-related or with non-sediment-related attributes higher overall performance when the model has been trained with the sediment-related attributes. The Whole-CONUS model captures the (imperfect) relationships between sediment-related attributes and sediment dynamics, exposing the spatial heterogeneity of SSC characteristics (higher overall performance from the Local-CONUS model compared to the Whole-CONUS model) that require more descriptors to overcome the interconnectedness and the heterogeneous characteristics. Future work should seek these critical inputs to improve the Whole-CONUS model, allowing us to simulate sediment at large scales with high accuracy.