diff --git a/dpdata/lammps/dump.py b/dpdata/lammps/dump.py
index fe549b95..3f998d1c 100644
--- a/dpdata/lammps/dump.py
+++ b/dpdata/lammps/dump.py
@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 from __future__ import annotations
 
+import itertools
 import os
 import sys
 from typing import TYPE_CHECKING
@@ -175,7 +176,177 @@ def box2dumpbox(orig, box):
     return bounds, tilt
 
 
-def load_file(fname: FileType, begin=0, step=1):
+def get_frame_nlines(fname: FileType):
+    """
+    Determine the number of lines per frame in a LAMMPS dump file.
+    
+    Parameters
+    ----------
+    fname : FileType
+        The dump file name
+        
+    Returns
+    -------
+    int
+        Number of lines per frame
+    """
+    with open_file(fname) as fp:
+        frame_start = None
+        line_count = 0
+        
+        while True:
+            line = fp.readline()
+            if not line:
+                break
+            line_count += 1
+            
+            if "ITEM: TIMESTEP" in line:
+                if frame_start is None:
+                    frame_start = line_count
+                else:
+                    # Found the start of the second frame
+                    return line_count - frame_start
+    
+    # If we only have one frame, return the total line count
+    return line_count
+
+
+def read_frames(fname: FileType, f_idx: list[int]):
+    """
+    Efficiently read only specified frames from a LAMMPS dump file.
+    
+    Parameters
+    ----------
+    fname : FileType
+        The dump file name
+    f_idx : list[int]
+        List of frame indices to read (0-based)
+        
+    Returns
+    -------
+    list[str]
+        Lines for the requested frames
+    """
+    if not f_idx:
+        return []
+    
+    # Sort frame indices for efficient sequential reading
+    sorted_indices = sorted(set(f_idx))
+    nlines = get_frame_nlines(fname)
+    
+    lines = []
+    with open_file(fname) as fp:
+        frame_idx = 0
+        target_idx = 0
+        
+        # Use itertools.zip_longest to read frames in blocks
+        while target_idx < len(sorted_indices):
+            # Read a frame block
+            frame_lines = []
+            for _ in range(nlines):
+                line = fp.readline()
+                if not line:
+                    return lines  # End of file
+                frame_lines.append(line.rstrip("\n"))
+            
+            # Check if this is a frame we want
+            if frame_idx == sorted_indices[target_idx]:
+                lines.extend(frame_lines)
+                target_idx += 1
+            
+            frame_idx += 1
+            
+            # Skip ahead if the next target frame is far away
+            if target_idx < len(sorted_indices):
+                frames_to_skip = sorted_indices[target_idx] - frame_idx
+                if frames_to_skip > 0:
+                    # Skip frames by reading and discarding lines
+                    for _ in range(frames_to_skip * nlines):
+                        line = fp.readline()
+                        if not line:
+                            return lines
+                    frame_idx += frames_to_skip
+    
+    return lines
+
+
+def load_frames_from_trajectories(frames_dict, **kwargs):
+    """
+    Load frames from multiple trajectory files efficiently.
+    
+    This implements the pattern described in the issue:
+    frames_dict = {
+      Trajectory0: [23, 56, 78],
+      Trajectory1: [22],
+      ...
+    }
+    
+    Parameters
+    ----------
+    frames_dict : dict
+        Dictionary mapping trajectory file paths to lists of frame indices
+    **kwargs
+        Additional arguments passed to system_data (e.g., type_map, unwrap, input_file)
+        
+    Returns
+    -------
+    dict
+        Combined system data from all requested frames
+    """
+    combined_data = None
+    
+    for traj_file, f_idx in frames_dict.items():
+        if not f_idx:
+            continue
+            
+        # Read specific frames from this trajectory
+        lines = read_frames(traj_file, f_idx)
+        if not lines:
+            continue
+            
+        # Convert to system data
+        data = system_data(lines, **kwargs)
+        
+        if combined_data is None:
+            combined_data = data.copy()
+        else:
+            # Append data from this trajectory
+            combined_data["cells"] = np.concatenate([combined_data["cells"], data["cells"]], axis=0)
+            combined_data["coords"] = np.concatenate([combined_data["coords"], data["coords"]], axis=0)
+            
+            if "spins" in combined_data and "spins" in data:
+                combined_data["spins"] = np.concatenate([combined_data["spins"], data["spins"]], axis=0)
+            elif "spins" in data:
+                combined_data["spins"] = data["spins"]
+    
+    return combined_data if combined_data is not None else {}
+
+
+def load_file(fname: FileType, begin=0, step=1, f_idx: list[int] = None):
+    """
+    Load frames from a LAMMPS dump file.
+    
+    Parameters
+    ----------
+    fname : FileType
+        The dump file name
+    begin : int, optional
+        The begin frame index (used when f_idx is None)
+    step : int, optional
+        The step between frames (used when f_idx is None)
+    f_idx : list[int], optional
+        Specific frame indices to load. If provided, begin and step are ignored.
+        
+    Returns
+    -------
+    list[str]
+        Lines for the requested frames
+    """
+    if f_idx is not None:
+        # Use efficient frame reading for specific indices
+        return read_frames(fname, f_idx)
+    
+    # Original implementation for begin/step reading
     lines = []
     buff = []
     cc = -1
diff --git a/dpdata/plugins/lammps.py b/dpdata/plugins/lammps.py
index c7e5c765..617f4fd8 100644
--- a/dpdata/plugins/lammps.py
+++ b/dpdata/plugins/lammps.py
@@ -69,6 +69,7 @@ def from_system(
         step: int = 1,
         unwrap: bool = False,
         input_file: str = None,
+        f_idx: list[int] = None,
         **kwargs,
     ):
         """Read the data from a lammps dump file.
@@ -87,13 +88,15 @@ def from_system(
             Whether to unwrap the coordinates
         input_file : str, optional
             The input file name
+        f_idx : list[int], optional
+            Specific frame indices to load. If provided, begin and step are ignored.
 
         Returns
         -------
         dict
             The system data
         """
-        lines = dpdata.lammps.dump.load_file(file_name, begin=begin, step=step)
+        lines = dpdata.lammps.dump.load_file(file_name, begin=begin, step=step, f_idx=f_idx)
         data = dpdata.lammps.dump.system_data(
             lines, type_map, unwrap=unwrap, input_file=input_file
         )
diff --git a/tests/test_lammps_dump_efficient_read.py b/tests/test_lammps_dump_efficient_read.py
new file mode 100644
index 00000000..899ec943
--- /dev/null
+++ b/tests/test_lammps_dump_efficient_read.py
@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""Test efficient frame reading functionality for LAMMPS dump files."""
+
+from __future__ import annotations
+
+import os
+import unittest
+
+import numpy as np
+from comp_sys import CompSys, IsPBC
+from context import dpdata
+import dpdata.lammps.dump as dump
+
+
+class TestLAMMPSDumpEfficientRead(unittest.TestCase, CompSys, IsPBC):
+    def setUp(self):
+        self.dump_file = os.path.join("poscars", "conf.dump")
+        self.type_map = ["O", "H"]
+        self.places = 6
+        self.e_places = 6
+        self.f_places = 6
+        self.v_places = 4
+        
+        # Set up comparison systems for inherited tests
+        # Use the new efficient method as system_1
+        self.system_1 = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map, f_idx=[0])
+        # Use traditional method as system_2
+        self.system_2 = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map).sub_system([0])
+
+    def test_get_frame_nlines(self):
+        """Test frame line count detection."""
+        nlines = dump.get_frame_nlines(self.dump_file)
+        self.assertEqual(nlines, 11)  # Expected based on file structure
+
+    def test_read_frames_single(self):
+        """Test reading a single frame."""
+        lines = dump.read_frames(self.dump_file, [1])
+        self.assertEqual(len(lines), 11)
+        self.assertTrue(lines[0].startswith("ITEM: TIMESTEP"))
+        self.assertEqual(lines[1], "1")  # Second frame has timestep 1
+
+    def test_read_frames_multiple(self):
+        """Test reading multiple frames."""
+        lines = dump.read_frames(self.dump_file, [0, 1])
+        self.assertEqual(len(lines), 22)  # 11 lines per frame * 2 frames
+
+    def test_read_frames_out_of_order(self):
+        """Test reading frames in non-sequential order."""
+        lines1 = dump.read_frames(self.dump_file, [1, 0])
+        lines2 = dump.read_frames(self.dump_file, [0, 1])
+        self.assertEqual(len(lines1), len(lines2))
+
+    def test_read_frames_empty(self):
+        """Test reading with empty frame list."""
+        lines = dump.read_frames(self.dump_file, [])
+        self.assertEqual(len(lines), 0)
+
+    def test_load_file_with_f_idx(self):
+        """Test enhanced load_file with f_idx parameter."""
+        # Load specific frame
+        lines = dump.load_file(self.dump_file, f_idx=[1])
+        self.assertEqual(len(lines), 11)
+        
+        # Load multiple frames
+        lines = dump.load_file(self.dump_file, f_idx=[0, 1])
+        self.assertEqual(len(lines), 22)
+        
+        # Test that f_idx overrides begin/step
+        lines = dump.load_file(self.dump_file, begin=1, step=1, f_idx=[0])
+        self.assertEqual(len(lines), 11)
+
+    def test_system_with_f_idx(self):
+        """Test dpdata.System with f_idx parameter."""
+        # Load all frames for comparison
+        system_all = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map)
+        
+        # Load only second frame
+        system_f1 = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map, f_idx=[1])
+        
+        self.assertEqual(len(system_all.data["coords"]), 2)
+        self.assertEqual(len(system_f1.data["coords"]), 1)
+        
+        # Check that the frame data matches
+        np.testing.assert_array_almost_equal(
+            system_all.data["coords"][1], 
+            system_f1.data["coords"][0]
+        )
+        np.testing.assert_array_almost_equal(
+            system_all.data["cells"][1], 
+            system_f1.data["cells"][0]
+        )
+
+    def test_load_frames_from_trajectories(self):
+        """Test the frames_dict pattern."""
+        frames_dict = {
+            self.dump_file: [0, 1]
+        }
+        
+        data = dump.load_frames_from_trajectories(frames_dict, type_map=self.type_map)
+        
+        self.assertIn("coords", data)
+        self.assertIn("cells", data)
+        self.assertEqual(len(data["coords"]), 2)
+        self.assertEqual(len(data["cells"]), 2)
+
+    def test_load_frames_from_trajectories_single(self):
+        """Test the frames_dict pattern with single frame."""
+        frames_dict = {
+            self.dump_file: [1]
+        }
+        
+        data = dump.load_frames_from_trajectories(frames_dict, type_map=self.type_map)
+        
+        self.assertIn("coords", data)
+        self.assertIn("cells", data)
+        self.assertEqual(len(data["coords"]), 1)
+        self.assertEqual(len(data["cells"]), 1)
+
+    def test_efficiency_comparison(self):
+        """Compare efficiency by verifying we get the same results."""
+        # Traditional approach: load all then filter
+        system_traditional = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map)
+        filtered_traditional = system_traditional.sub_system([1])
+        
+        # New efficient approach: load only frame 1
+        system_efficient = dpdata.System(self.dump_file, fmt="lammps/dump", type_map=self.type_map, f_idx=[1])
+        
+        # Results should be identical
+        np.testing.assert_array_almost_equal(
+            filtered_traditional.data["coords"][0], 
+            system_efficient.data["coords"][0]
+        )
+        np.testing.assert_array_almost_equal(
+            filtered_traditional.data["cells"][0], 
+            system_efficient.data["cells"][0]
+        )
+
+    def setUp_comp_sys(self):
+        """Set up comparison systems for inherited tests."""
+        pass  # Already set up in setUp
+
+
+if __name__ == "__main__":
+    unittest.main()
\ No newline at end of file